ggml : try to improve threading

* ggml : simplify the SIMD code

* ggml : generic reduce for all register sizes + comments

CMakeLists.txt

@@ -132,6 +132,12 @@ if (WHISPER_ALL_WARNINGS)
             -Wstrict-prototypes             \
             -Wpointer-arith                 \
         ")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} \
+            -Wall                           \
+            -Wextra                         \
+            -Wpedantic                      \
+            -Wcast-qual                     \
+        ")
     else()
         # todo : msvc
     endif()
@@ -178,7 +184,9 @@ endif()
 set(TARGET whisper)
 
 add_library(${TARGET}
+    ggml.h
     ggml.c
+    whisper.h
     whisper.cpp
     )
 

Makefile

@@ -105,6 +105,12 @@ endif
 ifeq ($(UNAME_M),amd64)
 	CFLAGS += -mavx -mavx2 -mfma -mf16c
 endif
+ifeq ($(UNAME_M),ppc64le)
+	POWER9_M := $(shell grep "POWER9" /proc/cpuinfo)
+	ifneq (,$(findstring POWER9,$(POWER9_M)))
+		CFLAGS += -mpower9-vector
+	endif
+endif
 ifndef WHISPER_NO_ACCELERATE
 	# Mac M1 - include Accelerate framework
 	ifeq ($(UNAME_S),Darwin)

README.md

@@ -447,12 +447,13 @@ or manually from here:
 For more details, see the conversion script [models/convert-pt-to-ggml.py](models/convert-pt-to-ggml.py) or the README
 in [models](models).
 
-## Bindings
+## [Bindings](https://github.com/ggerganov/whisper.cpp/discussions/categories/bindings)
 
-- [X] Rust: [tazz4843/whisper-rs](https://github.com/tazz4843/whisper-rs)
-- [X] Objective-C / Swift: [ggerganov/whisper.spm](https://github.com/ggerganov/whisper.spm)
-- [X] Javascript: [bindings/javascript](bindings/javascript)
-- [ ] Python: soon
+- [X] Rust: [tazz4843/whisper-rs](https://github.com/tazz4843/whisper-rs) | [#310](https://github.com/ggerganov/whisper.cpp/discussions/310)
+- [X] Javascript: [bindings/javascript](bindings/javascript) | [#309](https://github.com/ggerganov/whisper.cpp/discussions/309)
+- [X] Go: [bindings/go](bindings/go) | [#312](https://github.com/ggerganov/whisper.cpp/discussions/312)
+- [X] Objective-C / Swift: [ggerganov/whisper.spm](https://github.com/ggerganov/whisper.spm) | [#313](https://github.com/ggerganov/whisper.cpp/discussions/313)
+- [ ] Python: soon | [WIP](https://github.com/ggerganov/whisper.cpp/issues/9)
 
 ## Examples
 
@@ -467,6 +468,7 @@ Some of the examples are even ported to run in the browser using WebAssembly. Ch
 | [command](examples/command) | [command.wasm](examples/command.wasm) | Basic voice assistant example for receiving voice commands from the mic |
 | [talk](examples/talk) | [talk.wasm](examples/talk.wasm) | Talk with a GPT-2 bot |
 | [whisper.objc](examples/whisper.objc) | | iOS mobile application using whisper.cpp |
+| [whisper.swiftui](examples/whisper.swiftui) | | SwiftUI iOS / macOS application using whisper.cpp |
 | [whisper.android](examples/whisper.android) | | Android mobile application using whisper.cpp |
 | [whisper.nvim](examples/whisper.nvim) | | Speech-to-text plugin for Neovim |
 | [generate-karaoke.sh](examples/generate-karaoke.sh) | | Helper script to easily [generate a karaoke video](https://youtu.be/uj7hVta4blM) of raw audio capture |

bindings/go/.gitignore

@@ -0,0 +1,3 @@
+build
+models
+go.sum

bindings/go/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2022 David Thorpe
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

bindings/go/Makefile

@@ -0,0 +1,39 @@
+CMAKE := $(shell which cmake)
+BUILD_DIR := "build"
+MODELS_DIR := "models"
+EXAMPLES_DIR := $(wildcard examples/*)
+C_INCLUDE_PATH := "../.."
+
+all: clean whisper examples
+
+whisper: mkdir
+	@echo Build whisper
+	@${CMAKE} -S ../.. -B ${BUILD_DIR} -D BUILD_SHARED_LIBS=off -D WHISPER_NO_AVX2=on
+	@${CMAKE} --build ${BUILD_DIR} --target whisper
+
+test: model-small whisper modtidy
+	@go test -v .
+	@go test -v ./pkg/whisper/...
+
+examples: $(EXAMPLES_DIR)
+
+model-small: mkdir examples/go-model-download
+	@${BUILD_DIR}/go-model-download -out models small.en
+
+$(EXAMPLES_DIR): mkdir whisper modtidy
+	@echo Build example $(notdir $@)
+	@go build ${BUILD_FLAGS} -o ${BUILD_DIR}/$(notdir $@) ./$@
+
+mkdir:
+	@echo Mkdir ${BUILD_DIR}
+	@install -d ${BUILD_DIR}
+	@echo Mkdir ${MODELS_DIR}
+	@install -d ${MODELS_DIR}
+
+modtidy:
+	@go mod tidy
+
+clean: 
+	@echo Clean
+	@rm -fr $(BUILD_DIR)
+	@go clean

bindings/go/README.md

@@ -0,0 +1,77 @@
+# Go bindings for Whisper
+
+This package provides Go bindings for whisper.cpp. They have been tested on:
+
+  * Darwin (OS X) 12.6 on x64_64
+  * Debian Linux on arm64
+  * Fedora Linux on x86_64
+
+The "low level" bindings are in the `bindings/go` directory and there is a more
+Go-style package in the `bindings/go/pkg/whisper` directory. The most simple usage
+is as follows:
+
+```go
+import (
+	"github.com/ggerganov/whisper.cpp/bindings/go/pkg/whisper"
+)
+
+func main() {
+	var modelpath string // Path to the model
+	var samples []float32 // Samples to process
+
+	// Load the model
+	model, err := whisper.New(modelpath)
+	if err != nil {
+		panic(err)
+	}
+	defer model.Close()
+
+	// Process samples
+	context, err := model.NewContext()
+	if err != nil {
+		panic(err)
+	}
+	if err := context.Process(samples, nil); err != nil {
+		return err
+	}
+
+	// Print out the results
+	for {
+		segment, err := context.NextSegment()
+		if err != nil {
+			break
+		}
+		fmt.Printf("[%6s->%6s] %s\n", segment.Start, segment.End, segment.Text)
+	}
+}
+```
+
+## Building & Testing
+
+In order to build, you need to have the Go compiler installed. You can get it from [here](https://golang.org/dl/). Run the tests with:
+
+```bash
+git clone https://github.com/ggerganov/whisper.cpp.git
+cd whisper.cpp/bindings/go
+make test
+```
+
+This will compile a static `libwhisper.a` in a `build` folder, download a model file, then run the tests. To build the examples:
+
+```bash
+make examples
+```
+
+The examples are placed in the `build` directory. Once built, you can download all the models with the following command:
+
+```bash
+./build/go-model-download -out models
+```
+
+And you can then test a model against samples with the following command:
+
+```bash
+./build/go-whisper -model models/ggml-tiny.en.bin samples/jfk.wav 
+```
+
+

bindings/go/doc.go

@@ -0,0 +1,5 @@
+/*
+github.com/ggerganov/whisper.cpp/bindings/go
+provides a speech-to-text service bindings for the Go programming language.
+*/
+package whisper

bindings/go/examples/go-model-download/context.go

@@ -0,0 +1,30 @@
+package main
+
+import (
+	"context"
+	"os"
+	"os/signal"
+)
+
+// ContextForSignal returns a context object which is cancelled when a signal
+// is received. It returns nil if no signal parameter is provided
+func ContextForSignal(signals ...os.Signal) context.Context {
+	if len(signals) == 0 {
+		return nil
+	}
+
+	ch := make(chan os.Signal)
+	ctx, cancel := context.WithCancel(context.Background())
+
+	// Send message on channel when signal received
+	signal.Notify(ch, signals...)
+
+	// When any signal received, call cancel
+	go func() {
+		<-ch
+		cancel()
+	}()
+
+	// Return success
+	return ctx
+}

bindings/go/examples/go-model-download/main.go

@@ -0,0 +1,206 @@
+package main
+
+import (
+	"context"
+	"flag"
+	"fmt"
+	"io"
+	"net/http"
+	"net/url"
+	"os"
+	"path/filepath"
+	"syscall"
+	"time"
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// CONSTANTS
+
+const (
+	srcUrl        = "https://huggingface.co/"                           // The location of the models
+	srcPathPrefix = "/datasets/ggerganov/whisper.cpp/resolve/main/ggml" // Filename prefix
+	srcExt        = ".bin"                                              // Filename extension
+	bufSize       = 1024 * 64                                           // Size of the buffer used for downloading the model
+)
+
+var (
+	// The models which will be downloaded, if no model is specified as an argument
+	modelNames = []string{"tiny.en", "tiny", "base.en", "base", "small.en", "small", "medium.en", "medium", "large-v1", "large"}
+)
+
+var (
+	// The output folder. When not set, use current working directory.
+	flagOut = flag.String("out", "", "Output folder")
+
+	// HTTP timeout parameter - will timeout if takes longer than this to download a model
+	flagTimeout = flag.Duration("timeout", 30*time.Minute, "HTTP timeout")
+
+	// Quiet parameter - will not print progress if set
+	flagQuiet = flag.Bool("quiet", false, "Quiet mode")
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// MAIN
+
+func main() {
+	flag.Usage = func() {
+		name := filepath.Base(flag.CommandLine.Name())
+		fmt.Fprintf(flag.CommandLine.Output(), "Usage: %s [options] <model>\n\n", name)
+		flag.PrintDefaults()
+	}
+	flag.Parse()
+
+	// Get output path
+	out, err := GetOut()
+	if err != nil {
+		fmt.Fprintln(os.Stderr, "Error:", err)
+		os.Exit(-1)
+	}
+
+	// Create context which quits on SIGINT or SIGQUIT
+	ctx := ContextForSignal(os.Interrupt, syscall.SIGQUIT)
+
+	// Progress filehandle
+	progress := os.Stdout
+	if *flagQuiet {
+		progress, err = os.Open(os.DevNull)
+		if err != nil {
+			fmt.Fprintln(os.Stderr, "Error:", err)
+			os.Exit(-1)
+		}
+		defer progress.Close()
+	}
+
+	// Download models - exit on error or interrupt
+	for _, model := range GetModels() {
+		url, err := URLForModel(model)
+		if err != nil {
+			fmt.Fprintln(os.Stderr, "Error:", err)
+			continue
+		} else if path, err := Download(ctx, progress, url, out); err == nil || err == io.EOF {
+			continue
+		} else if err == context.Canceled {
+			os.Remove(path)
+			fmt.Fprintln(progress, "\nInterrupted")
+			break
+		} else if err == context.DeadlineExceeded {
+			os.Remove(path)
+			fmt.Fprintln(progress, "Timeout downloading model")
+			continue
+		} else {
+			os.Remove(path)
+			fmt.Fprintln(os.Stderr, "Error:", err)
+			break
+		}
+	}
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// PUBLIC METHODS
+
+// GetOut returns the path to the output directory
+func GetOut() (string, error) {
+	if *flagOut == "" {
+		return os.Getwd()
+	}
+	if info, err := os.Stat(*flagOut); err != nil {
+		return "", err
+	} else if !info.IsDir() {
+		return "", fmt.Errorf("not a directory: %s", info.Name())
+	} else {
+		return *flagOut, nil
+	}
+}
+
+// GetModels returns the list of models to download
+func GetModels() []string {
+	if flag.NArg() == 0 {
+		return modelNames
+	} else {
+		return flag.Args()
+	}
+}
+
+// URLForModel returns the URL for the given model on huggingface.co
+func URLForModel(model string) (string, error) {
+	url, err := url.Parse(srcUrl)
+	if err != nil {
+		return "", err
+	} else {
+		url.Path = srcPathPrefix + "-" + model + srcExt
+	}
+	return url.String(), nil
+}
+
+// Download downloads the model from the given URL to the given output directory
+func Download(ctx context.Context, p io.Writer, model, out string) (string, error) {
+	// Create HTTP client
+	client := http.Client{
+		Timeout: *flagTimeout,
+	}
+
+	// Initiate the download
+	req, err := http.NewRequest("GET", model, nil)
+	if err != nil {
+		return "", err
+	}
+	resp, err := client.Do(req)
+	if err != nil {
+		return "", err
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		return "", fmt.Errorf("%s: %s", model, resp.Status)
+	}
+
+	// If output file exists and is the same size as the model, skip
+	path := filepath.Join(out, filepath.Base(model))
+	if info, err := os.Stat(path); err == nil && info.Size() == resp.ContentLength {
+		fmt.Fprintln(p, "Skipping", model, "as it already exists")
+		return "", nil
+	}
+
+	// Create file
+	w, err := os.Create(path)
+	if err != nil {
+		return "", err
+	}
+	defer w.Close()
+
+	// Report
+	fmt.Fprintln(p, "Downloading", model, "to", out)
+
+	// Progressively download the model
+	data := make([]byte, bufSize)
+	count, pct := int64(0), int64(0)
+	ticker := time.NewTicker(5 * time.Second)
+	for {
+		select {
+		case <-ctx.Done():
+			// Cancelled, return error
+			return path, ctx.Err()
+		case <-ticker.C:
+			pct = DownloadReport(p, pct, count, resp.ContentLength)
+		default:
+			// Read body
+			n, err := resp.Body.Read(data)
+			if err != nil {
+				DownloadReport(p, pct, count, resp.ContentLength)
+				return path, err
+			} else if m, err := w.Write(data[:n]); err != nil {
+				return path, err
+			} else {
+				count += int64(m)
+			}
+		}
+	}
+}
+
+// Report periodically reports the download progress when percentage changes
+func DownloadReport(w io.Writer, pct, count, total int64) int64 {
+	pct_ := count * 100 / total
+	if pct_ > pct {
+		fmt.Fprintf(w, "  ...%d MB written (%d%%)\n", count/1e6, pct_)
+	}
+	return pct_
+}

bindings/go/examples/go-whisper/flags.go

@@ -0,0 +1,61 @@
+package main
+
+import (
+	"flag"
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// TYPES
+
+type Flags struct {
+	*flag.FlagSet
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// LIFECYCLE
+
+func NewFlags(name string, args []string) (*Flags, error) {
+	flags := &Flags{
+		FlagSet: flag.NewFlagSet(name, flag.ContinueOnError),
+	}
+
+	// Register the command line arguments
+	registerFlags(flags)
+
+	// Parse command line
+	if err := flags.Parse(args); err != nil {
+		return nil, err
+	}
+
+	// Return success
+	return flags, nil
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// PUBLIC METHODS
+
+func (flags *Flags) GetModel() string {
+	return flags.Lookup("model").Value.String()
+}
+
+func (flags *Flags) GetLanguage() string {
+	return flags.Lookup("language").Value.String()
+}
+
+func (flags *Flags) IsSpeedup() bool {
+	return flags.Lookup("speedup").Value.String() == "true"
+}
+
+func (flags *Flags) IsTokens() bool {
+	return flags.Lookup("tokens").Value.String() == "true"
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// PRIVATE METHODS
+
+func registerFlags(flag *Flags) {
+	flag.String("model", "", "Path to the model file")
+	flag.String("language", "", "Language")
+	flag.Bool("speedup", false, "Enable speedup")
+	flag.Bool("tokens", false, "Display tokens")
+}

bindings/go/examples/go-whisper/main.go

@@ -0,0 +1,44 @@
+package main
+
+import (
+	"flag"
+	"fmt"
+	"os"
+	"path/filepath"
+
+	// Packages
+	whisper "github.com/ggerganov/whisper.cpp/bindings/go/pkg/whisper"
+)
+
+func main() {
+	flags, err := NewFlags(filepath.Base(os.Args[0]), os.Args[1:])
+	if err == flag.ErrHelp {
+		os.Exit(0)
+	} else if err != nil {
+		fmt.Fprintln(os.Stderr, err)
+		os.Exit(1)
+	} else if flags.GetModel() == "" {
+		fmt.Fprintln(os.Stderr, "Use -model flag to specify which model file to use")
+		os.Exit(1)
+	} else if flags.NArg() == 0 {
+		fmt.Fprintln(os.Stderr, "No input files specified")
+		os.Exit(1)
+	}
+
+	// Load model
+	model, err := whisper.New(flags.GetModel())
+	if err != nil {
+		fmt.Fprintln(os.Stderr, err)
+		os.Exit(1)
+	}
+	defer model.Close()
+
+	// Process files
+	for _, filename := range flags.Args() {
+		fmt.Println("Processing", filename)
+		if err := Process(model, filename, flags.GetLanguage(), flags.IsSpeedup(), flags.IsTokens()); err != nil {
+			fmt.Fprintln(os.Stderr, err)
+			continue
+		}
+	}
+}

bindings/go/examples/go-whisper/process.go

@@ -0,0 +1,80 @@
+package main
+
+import (
+	"fmt"
+	"io"
+	"os"
+	"time"
+
+	// Package imports
+	whisper "github.com/ggerganov/whisper.cpp/bindings/go/pkg/whisper"
+	wav "github.com/go-audio/wav"
+)
+
+func Process(model whisper.Model, path string, lang string, speedup, tokens bool) error {
+	var data []float32
+
+	// Create processing context
+	context, err := model.NewContext()
+	if err != nil {
+		return err
+	}
+
+	// Open the file
+	fh, err := os.Open(path)
+	if err != nil {
+		return err
+	}
+	defer fh.Close()
+
+	// Decode the WAV file
+	dec := wav.NewDecoder(fh)
+	if buf, err := dec.FullPCMBuffer(); err != nil {
+		return err
+	} else if dec.SampleRate != whisper.SampleRate {
+		return fmt.Errorf("unsupported sample rate: %d", dec.SampleRate)
+	} else if dec.NumChans != 1 {
+		return fmt.Errorf("unsupported number of channels: %d", dec.NumChans)
+	} else {
+		data = buf.AsFloat32Buffer().Data
+	}
+
+	// Set the parameters
+	var cb whisper.SegmentCallback
+	if lang != "" {
+		if err := context.SetLanguage(lang); err != nil {
+			return err
+		}
+	}
+	if speedup {
+		context.SetSpeedup(true)
+	}
+	if tokens {
+		cb = func(segment whisper.Segment) {
+			fmt.Printf("%02d [%6s->%6s] ", segment.Num, segment.Start.Truncate(time.Millisecond), segment.End.Truncate(time.Millisecond))
+			for _, token := range segment.Tokens {
+				fmt.Printf("%q ", token.Text)
+			}
+			fmt.Println("")
+		}
+	}
+
+	// Process the data
+	if err := context.Process(data, cb); err != nil {
+		return err
+	}
+
+	// Print out the results
+	for {
+		segment, err := context.NextSegment()
+		if err == io.EOF {
+			break
+		} else if err != nil {
+			return err
+		}
+		fmt.Printf("[%6s->%6s] %s\n", segment.Start.Truncate(time.Millisecond), segment.End.Truncate(time.Millisecond), segment.Text)
+	}
+
+	// Return success
+	return nil
+}

bindings/go/go.mod

@@ -0,0 +1,16 @@
+module github.com/ggerganov/whisper.cpp/bindings/go
+
+go 1.19
+
+require (
+	github.com/go-audio/wav v1.1.0
+	github.com/stretchr/testify v1.8.1
+)
+
+require (
+	github.com/davecgh/go-spew v1.1.1 // indirect
+	github.com/go-audio/audio v1.0.0 // indirect
+	github.com/go-audio/riff v1.0.0 // indirect
+	github.com/pmezard/go-difflib v1.0.0 // indirect
+	gopkg.in/yaml.v3 v3.0.1 // indirect
+)

bindings/go/params.go

@@ -0,0 +1,134 @@
+package whisper
+
+// This file defines the whisper_token, whisper_token_data and whisper_full_params
+// structures, which are used by the whisper_full() function.
+
+import (
+	"fmt"
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// CGO
+
+/*
+#include <whisper.h>
+*/
+import "C"
+
+///////////////////////////////////////////////////////////////////////////////
+// PUBLIC METHODS
+
+func (p *Params) SetTranslate(v bool) {
+	p.translate = toBool(v)
+}
+
+func (p *Params) SetNoContext(v bool) {
+	p.no_context = toBool(v)
+}
+
+func (p *Params) SetSingleSegment(v bool) {
+	p.single_segment = toBool(v)
+}
+
+func (p *Params) SetPrintSpecial(v bool) {
+	p.print_special = toBool(v)
+}
+
+func (p *Params) SetPrintProgress(v bool) {
+	p.print_progress = toBool(v)
+}
+
+func (p *Params) SetPrintRealtime(v bool) {
+	p.print_realtime = toBool(v)
+}
+
+func (p *Params) SetPrintTimestamps(v bool) {
+	p.print_timestamps = toBool(v)
+}
+
+func (p *Params) SetSpeedup(v bool) {
+	p.speed_up = toBool(v)
+}
+
+func (p *Params) SetLanguage(lang int) error {
+	str := C.whisper_lang_str(C.int(lang))
+	if str == nil {
+		return ErrInvalidLanguage
+	} else {
+		p.language = str
+	}
+	return nil
+}
+
+func (p *Params) Language() int {
+	if p.language == nil {
+		return -1
+	}
+	return int(C.whisper_lang_id(p.language))
+}
+
+func (p *Params) SetThreads(threads int) {
+	p.n_threads = C.int(threads)
+}
+
+func (p *Params) SetOffset(offset_ms int) {
+	p.offset_ms = C.int(offset_ms)
+}
+
+func (p *Params) SetDuration(duration_ms int) {
+	p.duration_ms = C.int(duration_ms)
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// PRIVATE METHODS
+
+func toBool(v bool) C.bool {
+	if v {
+		return C.bool(true)
+	}
+	return C.bool(false)
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// STRINGIFY
+
+func (p *Params) String() string {
+	str := "<whisper.params"
+	str += fmt.Sprintf(" strategy=%v", p.strategy)
+	str += fmt.Sprintf(" n_threads=%d", p.n_threads)
+	if p.language != nil {
+		str += fmt.Sprintf(" language=%s", C.GoString(p.language))
+	}
+	str += fmt.Sprintf(" n_max_text_ctx=%d", p.n_max_text_ctx)
+	str += fmt.Sprintf(" offset_ms=%d", p.offset_ms)
+	str += fmt.Sprintf(" duration_ms=%d", p.duration_ms)
+	if p.translate {
+		str += " translate"
+	}
+	if p.no_context {
+		str += " no_context"
+	}
+	if p.single_segment {
+		str += " single_segment"
+	}
+	if p.print_special {
+		str += " print_special"
+	}
+	if p.print_progress {
+		str += " print_progress"
+	}
+	if p.print_realtime {
+		str += " print_realtime"
+	}
+	if p.print_timestamps {
+		str += " print_timestamps"
+	}
+	if p.token_timestamps {
+		str += " token_timestamps"
+	}
+	if p.speed_up {
+		str += " speed_up"
+	}
+
+	return str + ">"
+}

bindings/go/pkg/whisper/consts.go

@@ -0,0 +1,27 @@
+package whisper
+
+import (
+	"errors"
+
+	// Bindings
+	whisper "github.com/ggerganov/whisper.cpp/bindings/go"
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// ERRORS
+
+var (
+	ErrUnableToLoadModel   = errors.New("unable to load model")
+	ErrInternalAppError    = errors.New("internal application error")
+	ErrProcessingFailed    = errors.New("processing failed")
+	ErrUnsupportedLanguage = errors.New("unsupported language")
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// CONSTANTS
+
+// SampleRate is the sample rate of the audio data.
+const SampleRate = whisper.SampleRate
+
+// SampleBits is the number of bytes per sample.
+const SampleBits = whisper.SampleBits

bindings/go/pkg/whisper/context.go

@@ -0,0 +1,145 @@
+package whisper
+
+import (
+	"io"
+	"strings"
+	"time"
+
+	// Bindings
+	whisper "github.com/ggerganov/whisper.cpp/bindings/go"
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// TYPES
+
+type context struct {
+	n      int
+	model  *model
+	params whisper.Params
+}
+
+// Make sure context adheres to the interface
+var _ Context = (*context)(nil)
+
+///////////////////////////////////////////////////////////////////////////////
+// LIFECYCLE
+
+func NewContext(model *model, params whisper.Params) (Context, error) {
+	context := new(context)
+	context.model = model
+	context.params = params
+
+	// Return success
+	return context, nil
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// PUBLIC METHODS
+
+// Set the language to use for speech recognition.
+func (context *context) SetLanguage(lang string) error {
+	if context.model.ctx == nil {
+		return ErrInternalAppError
+	}
+	if id := context.model.ctx.Whisper_lang_id(lang); id < 0 {
+		return ErrUnsupportedLanguage
+	} else if err := context.params.SetLanguage(id); err != nil {
+		return err
+	}
+	// Return success
+	return nil
+}
+
+// Get language
+func (context *context) Language() string {
+	return whisper.Whisper_lang_str(context.params.Language())
+}
+
+// Set speedup flag
+func (context *context) SetSpeedup(v bool) {
+	context.params.SetSpeedup(v)
+}
+
+// Process new sample data and return any errors
+func (context *context) Process(data []float32, cb SegmentCallback) error {
+	if context.model.ctx == nil {
+		return ErrInternalAppError
+	}
+	// If the callback is defined then we force on single_segment mode
+	if cb != nil {
+		context.params.SetSingleSegment(true)
+	}
+
+	// We don't do parallel processing at the moment
+	processors := 0
+	if processors > 1 {
+		if err := context.model.ctx.Whisper_full_parallel(context.params, data, processors, nil, func(new int) {
+			if cb != nil {
+				num_segments := context.model.ctx.Whisper_full_n_segments()
+				s0 := num_segments - new
+				for i := s0; i < num_segments; i++ {
+					cb(toSegment(context.model.ctx, i))
+				}
+			}
+		}); err != nil {
+			return err
+		}
+	} else if err := context.model.ctx.Whisper_full(context.params, data, nil, func(new int) {
+		if cb != nil {
+			num_segments := context.model.ctx.Whisper_full_n_segments()
+			s0 := num_segments - new
+			for i := s0; i < num_segments; i++ {
+				cb(toSegment(context.model.ctx, i))
+			}
+		}
+	}); err != nil {
+		return err
+	}
+
+	// Return success
+	return nil
+}
+
+// Return the next segment of tokens
+func (context *context) NextSegment() (Segment, error) {
+	if context.model.ctx == nil {
+		return Segment{}, ErrInternalAppError
+	}
+	if context.n >= context.model.ctx.Whisper_full_n_segments() {
+		return Segment{}, io.EOF
+	}
+
+	// Populate result
+	result := toSegment(context.model.ctx, context.n)
+
+	// Increment the cursor
+	context.n++
+
+	// Return success
+	return result, nil
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// PRIVATE METHODS
+
+func toSegment(ctx *whisper.Context, n int) Segment {
+	return Segment{
+		Num:    n,
+		Text:   strings.TrimSpace(ctx.Whisper_full_get_segment_text(n)),
+		Start:  time.Duration(ctx.Whisper_full_get_segment_t0(n)) * time.Millisecond * 10,
+		End:    time.Duration(ctx.Whisper_full_get_segment_t1(n)) * time.Millisecond * 10,
+		Tokens: toTokens(ctx, n),
+	}
+}
+
+func toTokens(ctx *whisper.Context, n int) []Token {
+	result := make([]Token, ctx.Whisper_full_n_tokens(n))
+	for i := 0; i < len(result); i++ {
+		result[i] = Token{
+			Id:   int(ctx.Whisper_full_get_token_id(n, i)),
+			Text: strings.TrimSpace(ctx.Whisper_full_get_token_text(n, i)),
+			P:    ctx.Whisper_full_get_token_p(n, i),
+		}
+	}
+	return result
+}

bindings/go/pkg/whisper/context_test.go

@@ -0,0 +1,55 @@
+package whisper_test
+
+import (
+	"os"
+	"testing"
+
+	// Packages
+	whisper "github.com/ggerganov/whisper.cpp/bindings/go/pkg/whisper"
+	assert "github.com/stretchr/testify/assert"
+)
+
+const (
+	ModelPath  = "../../models/ggml-tiny.bin"
+	SamplePath = "../../samples/jfk.wav"
+)
+
+func Test_Whisper_000(t *testing.T) {
+	assert := assert.New(t)
+	if _, err := os.Stat(ModelPath); os.IsNotExist(err) {
+		t.Skip("Skipping test, model not found:", ModelPath)
+	}
+	if _, err := os.Stat(SamplePath); os.IsNotExist(err) {
+		t.Skip("Skipping test, sample not found:", SamplePath)
+	}
+
+	// Load model
+	model, err := whisper.New(ModelPath)
+	assert.NoError(err)
+	assert.NotNil(model)
+	assert.NoError(model.Close())
+
+	t.Log("languages=", model.Languages())
+}
+
+func Test_Whisper_001(t *testing.T) {
+	assert := assert.New(t)
+	if _, err := os.Stat(ModelPath); os.IsNotExist(err) {
+		t.Skip("Skipping test, model not found:", ModelPath)
+	}
+	if _, err := os.Stat(SamplePath); os.IsNotExist(err) {
+		t.Skip("Skipping test, sample not found:", SamplePath)
+	}
+
+	// Load model
+	model, err := whisper.New(ModelPath)
+	assert.NoError(err)
+	assert.NotNil(model)
+	defer model.Close()
+
+	// Get context for decoding
+	ctx, err := model.NewContext()
+	assert.NoError(err)
+	assert.NotNil(ctx)
+
+}

bindings/go/pkg/whisper/doc.go

@@ -0,0 +1,4 @@
+/*
+This is the higher-level speech-to-text whisper.cpp API for go
+*/
+package whisper

bindings/go/pkg/whisper/interface.go

@@ -0,0 +1,63 @@
+package whisper
+
+import (
+	"io"
+	"time"
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// TYPES
+
+// SegmentCallback is the callback function for processing segments in real
+// time. It is called during the Process function
+type SegmentCallback func(Segment)
+
+// Model is the interface to a whisper model. Create a new model with the
+// function whisper.New(string)
+type Model interface {
+	io.Closer
+
+	// Return a new speech-to-text context.
+	NewContext() (Context, error)
+
+	// Return all languages supported.
+	Languages() []string
+}
+
+// Context is the speach recognition context.
+type Context interface {
+	SetLanguage(string) error // Set the language to use for speech recognition.
+	Language() string         // Get language
+	SetSpeedup(bool)          // Set speedup flag
+
+	// Process mono audio data and return any errors.
+	// If defined, newly generated segments are passed to the
+	// callback function during processing.
+	Process([]float32, SegmentCallback) error
+
+	// After process is called, return segments until the end of the stream
+	// is reached, when io.EOF is returned.
+	NextSegment() (Segment, error)
+}
+
+// Segment is the text result of a speech recognition.
+type Segment struct {
+	// Segment Number
+	Num int
+
+	// Time beginning and end timestamps for the segment.
+	Start, End time.Duration
+
+	// The text of the segment.
+	Text string
+
+	// The tokens of the segment.
+	Tokens []Token
+}
+
+// Token is a text or special token
+type Token struct {
+	Id   int
+	Text string
+	P    float32
+}

bindings/go/pkg/whisper/model.go

@@ -0,0 +1,95 @@
+package whisper
+
+import (
+	"fmt"
+	"os"
+	"runtime"
+
+	// Bindings
+	whisper "github.com/ggerganov/whisper.cpp/bindings/go"
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// TYPES
+
+type model struct {
+	path string
+	ctx  *whisper.Context
+}
+
+// Make sure model adheres to the interface
+var _ Model = (*model)(nil)
+
+///////////////////////////////////////////////////////////////////////////////
+// LIFECYCLE
+
+func New(path string) (*model, error) {
+	model := new(model)
+	if _, err := os.Stat(path); err != nil {
+		return nil, err
+	} else if ctx := whisper.Whisper_init(path); ctx == nil {
+		return nil, ErrUnableToLoadModel
+	} else {
+		model.ctx = ctx
+		model.path = path
+	}
+
+	// Return success
+	return model, nil
+}
+
+func (model *model) Close() error {
+	if model.ctx != nil {
+		model.ctx.Whisper_free()
+	}
+
+	// Release resources
+	model.ctx = nil
+
+	// Return success
+	return nil
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// STRINGIFY
+
+func (model *model) String() string {
+	str := "<whisper.model"
+	if model.ctx != nil {
+		str += fmt.Sprintf(" model=%q", model.path)
+	}
+	return str + ">"
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// PUBLIC METHODS
+
+// Return all recognized languages. Initially it is set to auto-detect
+func (model *model) Languages() []string {
+	result := make([]string, 0, whisper.Whisper_lang_max_id())
+	for i := 0; i < whisper.Whisper_lang_max_id(); i++ {
+		str := whisper.Whisper_lang_str(i)
+		if model.ctx.Whisper_lang_id(str) >= 0 {
+			result = append(result, str)
+		}
+	}
+	return result
+}
+
+func (model *model) NewContext() (Context, error) {
+	if model.ctx == nil {
+		return nil, ErrInternalAppError
+	}
+
+	// Create new context
+	params := model.ctx.Whisper_full_default_params(whisper.SAMPLING_GREEDY)
+	params.SetTranslate(false)
+	params.SetPrintSpecial(false)
+	params.SetPrintProgress(false)
+	params.SetPrintRealtime(false)
+	params.SetPrintTimestamps(false)
+	params.SetThreads(runtime.NumCPU())
+
+	// Return new context
+	return NewContext(model, params)
+}

bindings/go/whisper.go

@@ -0,0 +1,412 @@
+package whisper
+
+import (
+	"errors"
+	"unsafe"
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// CGO
+
+/*
+#cgo CFLAGS: -I${SRCDIR}/../..
+#cgo LDFLAGS: -L${SRCDIR}/build -lwhisper -lm -lstdc++
+#cgo darwin LDFLAGS: -framework Accelerate
+#include <whisper.h>
+#include <stdlib.h>
+
+extern void callNewSegment(void* user_data, int new);
+extern bool callEncoderBegin(void* user_data);
+
+// Text segment callback
+// Called on every newly generated text segment
+// Use the whisper_full_...() functions to obtain the text segments
+static void whisper_new_segment_cb(struct whisper_context* ctx, int n_new, void* user_data) {
+    if(user_data != NULL && ctx != NULL) {
+        callNewSegment(user_data, n_new);
+    }
+}
+
+// Encoder begin callback
+// If not NULL, called before the encoder starts
+// If it returns false, the computation is aborted
+static bool whisper_encoder_begin_cb(struct whisper_context* ctx, void* user_data) {
+    if(user_data != NULL && ctx != NULL) {
+        return callEncoderBegin(user_data);
+    }
+    return false;
+}
+
+// Get default parameters and set callbacks
+static struct whisper_full_params whisper_full_default_params_cb(struct whisper_context* ctx, enum whisper_sampling_strategy strategy) {
+	struct whisper_full_params params = whisper_full_default_params(strategy);
+	params.new_segment_callback = whisper_new_segment_cb;
+	params.new_segment_callback_user_data = (void*)(ctx);
+	params.encoder_begin_callback = whisper_encoder_begin_cb;
+	params.encoder_begin_callback_user_data = (void*)(ctx);
+	return params;
+}
+*/
+import "C"
+
+///////////////////////////////////////////////////////////////////////////////
+// TYPES
+
+type (
+	Context          C.struct_whisper_context
+	Token            C.whisper_token
+	TokenData        C.struct_whisper_token_data
+	SamplingStrategy C.enum_whisper_sampling_strategy
+	Params           C.struct_whisper_full_params
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// GLOBALS
+
+const (
+	SAMPLING_GREEDY      SamplingStrategy = C.WHISPER_SAMPLING_GREEDY
+	SAMPLING_BEAM_SEARCH SamplingStrategy = C.WHISPER_SAMPLING_BEAM_SEARCH
+)
+
+const (
+	SampleRate = C.WHISPER_SAMPLE_RATE                 // Expected sample rate, samples per second
+	SampleBits = uint16(unsafe.Sizeof(C.float(0))) * 8 // Sample size in bits
+	NumFFT     = C.WHISPER_N_FFT
+	NumMEL     = C.WHISPER_N_MEL
+	HopLength  = C.WHISPER_HOP_LENGTH
+	ChunkSize  = C.WHISPER_CHUNK_SIZE
+)
+
+var (
+	ErrTokenizerFailed  = errors.New("whisper_tokenize failed")
+	ErrAutoDetectFailed = errors.New("whisper_lang_auto_detect failed")
+	ErrConversionFailed = errors.New("whisper_convert failed")
+	ErrInvalidLanguage  = errors.New("invalid language")
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// PUBLIC METHODS
+
+// Allocates all memory needed for the model and loads the model from the given file.
+// Returns NULL on failure.
+func Whisper_init(path string) *Context {
+	cPath := C.CString(path)
+	defer C.free(unsafe.Pointer(cPath))
+	if ctx := C.whisper_init(cPath); ctx != nil {
+		return (*Context)(ctx)
+	} else {
+		return nil
+	}
+}
+
+// Frees all memory allocated by the model.
+func (ctx *Context) Whisper_free() {
+	C.whisper_free((*C.struct_whisper_context)(ctx))
+}
+
+// Convert RAW PCM audio to log mel spectrogram.
+// The resulting spectrogram is stored inside the provided whisper context.
+func (ctx *Context) Whisper_pcm_to_mel(data []float32, threads int) error {
+	if C.whisper_pcm_to_mel((*C.struct_whisper_context)(ctx), (*C.float)(&data[0]), C.int(len(data)), C.int(threads)) == 0 {
+		return nil
+	} else {
+		return ErrConversionFailed
+	}
+}
+
+// This can be used to set a custom log mel spectrogram inside the provided whisper context.
+// Use this instead of whisper_pcm_to_mel() if you want to provide your own log mel spectrogram.
+// n_mel must be 80
+func (ctx *Context) Whisper_set_mel(data []float32, n_mel int) error {
+	if C.whisper_set_mel((*C.struct_whisper_context)(ctx), (*C.float)(&data[0]), C.int(len(data)), C.int(n_mel)) == 0 {
+		return nil
+	} else {
+		return ErrConversionFailed
+	}
+}
+
+// Run the Whisper encoder on the log mel spectrogram stored inside the provided whisper context.
+// Make sure to call whisper_pcm_to_mel() or whisper_set_mel() first.
+// offset can be used to specify the offset of the first frame in the spectrogram.
+func (ctx *Context) Whisper_encode(offset, threads int) error {
+	if C.whisper_encode((*C.struct_whisper_context)(ctx), C.int(offset), C.int(threads)) == 0 {
+		return nil
+	} else {
+		return ErrConversionFailed
+	}
+}
+
+// Run the Whisper decoder to obtain the logits and probabilities for the next token.
+// Make sure to call whisper_encode() first.
+// tokens + n_tokens is the provided context for the decoder.
+// n_past is the number of tokens to use from previous decoder calls.
+func (ctx *Context) Whisper_decode(tokens []Token, past, threads int) error {
+	if C.whisper_decode((*C.struct_whisper_context)(ctx), (*C.whisper_token)(&tokens[0]), C.int(len(tokens)), C.int(past), C.int(threads)) == 0 {
+		return nil
+	} else {
+		return ErrConversionFailed
+	}
+}
+
+// whisper_sample_best() returns the token with the highest probability
+func (ctx *Context) Whisper_sample_best() TokenData {
+	return TokenData(C.whisper_sample_best((*C.struct_whisper_context)(ctx)))
+}
+
+// whisper_sample_timestamp() returns the most probable timestamp token
+func (ctx *Context) Whisper_sample_timestamp(is_initial bool) TokenData {
+	return TokenData(C.whisper_sample_timestamp((*C.struct_whisper_context)(ctx), C.bool(is_initial)))
+}
+
+// Convert the provided text into tokens. The tokens pointer must be large enough to hold the resulting tokens.
+// Returns the number of tokens on success
+func (ctx *Context) Whisper_tokenize(text string, tokens []Token) (int, error) {
+	cText := C.CString(text)
+	defer C.free(unsafe.Pointer(cText))
+	if n := C.whisper_tokenize((*C.struct_whisper_context)(ctx), cText, (*C.whisper_token)(&tokens[0]), C.int(len(tokens))); n >= 0 {
+		return int(n), nil
+	} else {
+		return 0, ErrTokenizerFailed
+	}
+}
+
+// Return the id of the specified language, returns -1 if not found
+func (ctx *Context) Whisper_lang_id(lang string) int {
+	return int(C.whisper_lang_id(C.CString(lang)))
+}
+
+// Largest language id (i.e. number of available languages - 1)
+func Whisper_lang_max_id() int {
+	return int(C.whisper_lang_max_id())
+}
+
+// Return the short string of the specified language id (e.g. 2 -> "de"),
+// returns empty string if not found
+func Whisper_lang_str(id int) string {
+	return C.GoString(C.whisper_lang_str(C.int(id)))
+}
+
+// Use mel data at offset_ms to try and auto-detect the spoken language
+// Make sure to call whisper_pcm_to_mel() or whisper_set_mel() first.
+// Returns the probabilities of all languages.
+// ref: https://github.com/openai/whisper/blob/main/whisper/decoding.py#L18-L69
+func (ctx *Context) Whisper_lang_auto_detect(offset_ms, n_threads int) ([]float32, error) {
+	probs := make([]float32, Whisper_lang_max_id()+1)
+	if n := int(C.whisper_lang_auto_detect((*C.struct_whisper_context)(ctx), C.int(offset_ms), C.int(n_threads), (*C.float)(&probs[0]))); n < 0 {
+		return nil, ErrAutoDetectFailed
+	} else {
+		return probs, nil
+	}
+}
+
+func (ctx *Context) Whisper_n_len() int {
+	return int(C.whisper_n_len((*C.struct_whisper_context)(ctx)))
+}
+
+func (ctx *Context) Whisper_n_vocab() int {
+	return int(C.whisper_n_vocab((*C.struct_whisper_context)(ctx)))
+}
+
+func (ctx *Context) Whisper_n_text_ctx() int {
+	return int(C.whisper_n_text_ctx((*C.struct_whisper_context)(ctx)))
+}
+
+func (ctx *Context) Whisper_is_multilingual() int {
+	return int(C.whisper_is_multilingual((*C.struct_whisper_context)(ctx)))
+}
+
+// The probabilities for the next token
+//func (ctx *Whisper_context) Whisper_get_probs() []float32 {
+//	return (*[1 << 30]float32)(unsafe.Pointer(C.whisper_get_probs((*C.struct_whisper_context)(ctx))))[:ctx.Whisper_n_vocab()]
+//}
+
+// Token Id -> String. Uses the vocabulary in the provided context
+func (ctx *Context) Whisper_token_to_str(token Token) string {
+	return C.GoString(C.whisper_token_to_str((*C.struct_whisper_context)(ctx), C.whisper_token(token)))
+}
+
+// Special tokens
+func (ctx *Context) Whisper_token_eot() Token {
+	return Token(C.whisper_token_eot((*C.struct_whisper_context)(ctx)))
+}
+
+// Special tokens
+func (ctx *Context) Whisper_token_sot() Token {
+	return Token(C.whisper_token_sot((*C.struct_whisper_context)(ctx)))
+}
+
+// Special tokens
+func (ctx *Context) Whisper_token_prev() Token {
+	return Token(C.whisper_token_prev((*C.struct_whisper_context)(ctx)))
+}
+
+// Special tokens
+func (ctx *Context) Whisper_token_solm() Token {
+	return Token(C.whisper_token_solm((*C.struct_whisper_context)(ctx)))
+}
+
+// Special tokens
+func (ctx *Context) Whisper_token_not() Token {
+	return Token(C.whisper_token_not((*C.struct_whisper_context)(ctx)))
+}
+
+// Special tokens
+func (ctx *Context) Whisper_token_beg() Token {
+	return Token(C.whisper_token_beg((*C.struct_whisper_context)(ctx)))
+}
+
+// Special tokens
+func (ctx *Context) Whisper_token_lang(lang_id int) Token {
+	return Token(C.whisper_token_lang((*C.struct_whisper_context)(ctx), C.int(lang_id)))
+}
+
+// Task tokens
+func Whisper_token_translate() Token {
+	return Token(C.whisper_token_translate())
+}
+
+// Task tokens
+func Whisper_token_transcribe() Token {
+	return Token(C.whisper_token_transcribe())
+}
+
+// Performance information
+func (ctx *Context) Whisper_print_timings() {
+	C.whisper_print_timings((*C.struct_whisper_context)(ctx))
+}
+
+// Performance information
+func (ctx *Context) Whisper_reset_timings() {
+	C.whisper_reset_timings((*C.struct_whisper_context)(ctx))
+}
+
+// Print system information
+func Whisper_print_system_info() string {
+	return C.GoString(C.whisper_print_system_info())
+}
+
+// Return default parameters for a strategy
+func (ctx *Context) Whisper_full_default_params(strategy SamplingStrategy) Params {
+	// Get default parameters
+	return Params(C.whisper_full_default_params_cb((*C.struct_whisper_context)(ctx), C.enum_whisper_sampling_strategy(strategy)))
+}
+
+// Run the entire model: PCM -> log mel spectrogram -> encoder -> decoder -> text
+// Uses the specified decoding strategy to obtain the text.
+func (ctx *Context) Whisper_full(params Params, samples []float32, encoderBeginCallback func() bool, newSegmentCallback func(int)) error {
+	registerEncoderBeginCallback(ctx, encoderBeginCallback)
+	registerNewSegmentCallback(ctx, newSegmentCallback)
+	defer registerEncoderBeginCallback(ctx, nil)
+	defer registerNewSegmentCallback(ctx, nil)
+	if C.whisper_full((*C.struct_whisper_context)(ctx), (C.struct_whisper_full_params)(params), (*C.float)(&samples[0]), C.int(len(samples))) == 0 {
+		return nil
+	} else {
+		return ErrConversionFailed
+	}
+}
+
+// Split the input audio in chunks and process each chunk separately using whisper_full()
+// It seems this approach can offer some speedup in some cases.
+// However, the transcription accuracy can be worse at the beginning and end of each chunk.
+func (ctx *Context) Whisper_full_parallel(params Params, samples []float32, processors int, encoderBeginCallback func() bool, newSegmentCallback func(int)) error {
+	registerEncoderBeginCallback(ctx, encoderBeginCallback)
+	registerNewSegmentCallback(ctx, newSegmentCallback)
+	defer registerEncoderBeginCallback(ctx, nil)
+	defer registerNewSegmentCallback(ctx, nil)
+
+	if C.whisper_full_parallel((*C.struct_whisper_context)(ctx), (C.struct_whisper_full_params)(params), (*C.float)(&samples[0]), C.int(len(samples)), C.int(processors)) == 0 {
+		return nil
+	} else {
+		return ErrConversionFailed
+	}
+}
+
+// Number of generated text segments.
+// A segment can be a few words, a sentence, or even a paragraph.
+func (ctx *Context) Whisper_full_n_segments() int {
+	return int(C.whisper_full_n_segments((*C.struct_whisper_context)(ctx)))
+}
+
+// Get the start and end time of the specified segment.
+func (ctx *Context) Whisper_full_get_segment_t0(segment int) int64 {
+	return int64(C.whisper_full_get_segment_t0((*C.struct_whisper_context)(ctx), C.int(segment)))
+}
+
+// Get the start and end time of the specified segment.
+func (ctx *Context) Whisper_full_get_segment_t1(segment int) int64 {
+	return int64(C.whisper_full_get_segment_t1((*C.struct_whisper_context)(ctx), C.int(segment)))
+}
+
+// Get the text of the specified segment.
+func (ctx *Context) Whisper_full_get_segment_text(segment int) string {
+	return C.GoString(C.whisper_full_get_segment_text((*C.struct_whisper_context)(ctx), C.int(segment)))
+}
+
+// Get number of tokens in the specified segment.
+func (ctx *Context) Whisper_full_n_tokens(segment int) int {
+	return int(C.whisper_full_n_tokens((*C.struct_whisper_context)(ctx), C.int(segment)))
+}
+
+// Get the token text of the specified token index in the specified segment.
+func (ctx *Context) Whisper_full_get_token_text(segment int, token int) string {
+	return C.GoString(C.whisper_full_get_token_text((*C.struct_whisper_context)(ctx), C.int(segment), C.int(token)))
+}
+
+// Get the token of the specified token index in the specified segment.
+func (ctx *Context) Whisper_full_get_token_id(segment int, token int) Token {
+	return Token(C.whisper_full_get_token_id((*C.struct_whisper_context)(ctx), C.int(segment), C.int(token)))
+}
+
+// Get token data for the specified token in the specified segment.
+// This contains probabilities, timestamps, etc.
+func (ctx *Context) whisper_full_get_token_data(segment int, token int) TokenData {
+	return TokenData(C.whisper_full_get_token_data((*C.struct_whisper_context)(ctx), C.int(segment), C.int(token)))
+}
+
+// Get the probability of the specified token in the specified segment.
+func (ctx *Context) Whisper_full_get_token_p(segment int, token int) float32 {
+	return float32(C.whisper_full_get_token_p((*C.struct_whisper_context)(ctx), C.int(segment), C.int(token)))
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// CALLBACKS
+
+var (
+	cbNewSegment   = make(map[unsafe.Pointer]func(int))
+	cbEncoderBegin = make(map[unsafe.Pointer]func() bool)
+)
+
+func registerNewSegmentCallback(ctx *Context, fn func(int)) {
+	if fn == nil {
+		delete(cbNewSegment, unsafe.Pointer(ctx))
+	} else {
+		cbNewSegment[unsafe.Pointer(ctx)] = fn
+	}
+}
+
+func registerEncoderBeginCallback(ctx *Context, fn func() bool) {
+	if fn == nil {
+		delete(cbEncoderBegin, unsafe.Pointer(ctx))
+	} else {
+		cbEncoderBegin[unsafe.Pointer(ctx)] = fn
+	}
+}
+
+//export callNewSegment
+func callNewSegment(user_data unsafe.Pointer, new C.int) {
+	if fn, ok := cbNewSegment[user_data]; ok {
+		fn(int(new))
+	}
+}
+
+//export callEncoderBegin
+func callEncoderBegin(user_data unsafe.Pointer) C.bool {
+	if fn, ok := cbEncoderBegin[user_data]; ok {
+		if fn() {
+			return C.bool(true)
+		} else {
+			return C.bool(false)
+		}
+	}
+	return true
+}

bindings/go/whisper_test.go

@@ -0,0 +1,110 @@
+package whisper_test
+
+import (
+	"os"
+	"runtime"
+	"testing"
+	"time"
+
+	// Packages
+	whisper "github.com/ggerganov/whisper.cpp/bindings/go"
+	wav "github.com/go-audio/wav"
+	assert "github.com/stretchr/testify/assert"
+)
+
+const (
+	ModelPath  = "models/ggml-small.en.bin"
+	SamplePath = "samples/jfk.wav"
+)
+
+func Test_Whisper_000(t *testing.T) {
+	assert := assert.New(t)
+	if _, err := os.Stat(ModelPath); os.IsNotExist(err) {
+		t.Skip("Skipping test, model not found:", ModelPath)
+	}
+	ctx := whisper.Whisper_init(ModelPath)
+	assert.NotNil(ctx)
+	ctx.Whisper_free()
+}
+
+func Test_Whisper_001(t *testing.T) {
+	assert := assert.New(t)
+	if _, err := os.Stat(ModelPath); os.IsNotExist(err) {
+		t.Skip("Skipping test, model not found:", ModelPath)
+	}
+	if _, err := os.Stat(SamplePath); os.IsNotExist(err) {
+		t.Skip("Skipping test, sample not found:", SamplePath)
+	}
+
+	// Open samples
+	fh, err := os.Open(SamplePath)
+	assert.NoError(err)
+	defer fh.Close()
+
+	// Read samples
+	d := wav.NewDecoder(fh)
+	buf, err := d.FullPCMBuffer()
+	assert.NoError(err)
+
+	// Run whisper
+	ctx := whisper.Whisper_init(ModelPath)
+	assert.NotNil(ctx)
+	defer ctx.Whisper_free()
+	assert.NoError(ctx.Whisper_full(ctx.Whisper_full_default_params(whisper.SAMPLING_GREEDY), buf.AsFloat32Buffer().Data, nil, nil))
+
+	// Print out tokens
+	num_segments := ctx.Whisper_full_n_segments()
+	assert.GreaterOrEqual(num_segments, 1)
+	for i := 0; i < num_segments; i++ {
+		str := ctx.Whisper_full_get_segment_text(i)
+		assert.NotEmpty(str)
+		t0 := time.Duration(ctx.Whisper_full_get_segment_t0(i)) * time.Millisecond
+		t1 := time.Duration(ctx.Whisper_full_get_segment_t1(i)) * time.Millisecond
+		t.Logf("[%6s->%-6s] %q", t0, t1, str)
+	}
+}
+
+func Test_Whisper_002(t *testing.T) {
+	assert := assert.New(t)
+	for i := 0; i < whisper.Whisper_lang_max_id(); i++ {
+		str := whisper.Whisper_lang_str(i)
+		assert.NotEmpty(str)
+		t.Log(str)
+	}
+}
+
+func Test_Whisper_003(t *testing.T) {
+	threads := runtime.NumCPU()
+	assert := assert.New(t)
+	if _, err := os.Stat(ModelPath); os.IsNotExist(err) {
+		t.Skip("Skipping test, model not found:", ModelPath)
+	}
+	if _, err := os.Stat(SamplePath); os.IsNotExist(err) {
+		t.Skip("Skipping test, sample not found:", SamplePath)
+	}
+
+	// Open samples
+	fh, err := os.Open(SamplePath)
+	assert.NoError(err)
+	defer fh.Close()
+
+	// Read samples
+	d := wav.NewDecoder(fh)
+	buf, err := d.FullPCMBuffer()
+	assert.NoError(err)
+
+	// Make the model
+	ctx := whisper.Whisper_init(ModelPath)
+	assert.NotNil(ctx)
+	defer ctx.Whisper_free()
+
+	// Get MEL
+	assert.NoError(ctx.Whisper_pcm_to_mel(buf.AsFloat32Buffer().Data, threads))
+
+	// Get Languages
+	languages, err := ctx.Whisper_lang_auto_detect(0, threads)
+	assert.NoError(err)
+	for i, p := range languages {
+		t.Logf("%s: %f", whisper.Whisper_lang_str(i), p)
+	}
+}

examples/bench/bench.cpp

@@ -33,7 +33,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     return true;
 }
 
-void whisper_print_usage(int argc, char ** argv, const whisper_params & params) {
+void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
     fprintf(stderr, "\n");
     fprintf(stderr, "usage: %s [options]\n", argv[0]);
     fprintf(stderr, "\n");

examples/command/command.cpp

@@ -41,8 +41,8 @@ struct whisper_params {
 
     std::string language  = "en";
     std::string model     = "models/ggml-base.en.bin";
-    std::string fname_out = "";
-    std::string commands  = "";
+    std::string fname_out;
+    std::string commands;
 };
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@@ -81,7 +81,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     return true;
 }
 
-void whisper_print_usage(int argc, char ** argv, const whisper_params & params) {
+void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
     fprintf(stderr, "\n");
     fprintf(stderr, "usage: %s [options]\n", argv[0]);
     fprintf(stderr, "\n");
@@ -387,7 +387,7 @@ bool vad_simple(std::vector<float> & pcmf32, int sample_rate, int last_ms, float
     float energy_all  = 0.0f;
     float energy_last = 0.0f;
 
-    for (size_t i = 0; i < n_samples; i++) {
+    for (int i = 0; i < n_samples; i++) {
         energy_all += fabsf(pcmf32[i]);
 
         if (i >= n_samples - n_samples_last) {
@@ -576,10 +576,10 @@ int main(int argc, char ** argv) {
     std::vector<std::string> allowed_commands;
     std::vector<std::vector<whisper_token>> allowed_tokens;
 
-    std::string k_prompt = "";
+    std::string k_prompt;
     std::vector<whisper_token> k_tokens;
 
-    if (params.commands != "") {
+    if (!params.commands.empty()) {
         fprintf(stderr, "\n");
         fprintf(stderr, "%s: guided mode\n", __func__);
 
@@ -594,7 +594,7 @@ int main(int argc, char ** argv) {
             whisper_token tokens[1024];
             allowed_tokens.emplace_back();
 
-            for (int l = 0; l < cmd.size(); ++l) {
+            for (int l = 0; l < (int) cmd.size(); ++l) {
                 // NOTE: very important to add the whitespace !
                 //       the reason is that the first decoded token starts with a whitespace too!
                 std::string ss = std::string(" ") + cmd.substr(0, l + 1);
@@ -808,7 +808,7 @@ int main(int argc, char ** argv) {
 
                 double psum = 0.0;
                 for (int i = 0; i < (int) allowed_commands.size(); ++i) {
-                    probs_id.push_back(std::make_pair(probs[allowed_tokens[i][0]], i));
+                    probs_id.emplace_back(probs[allowed_tokens[i][0]], i);
                     for (int j = 1; j < (int) allowed_tokens[i].size(); ++j) {
                         probs_id.back().first += probs[allowed_tokens[i][j]];
                     }
@@ -843,15 +843,15 @@ int main(int argc, char ** argv) {
 
                 // best command
                 {
+                    const auto t_end = std::chrono::high_resolution_clock::now();
+
                     fprintf(stdout, "\n");
                     fprintf(stdout, "%s: detected command: %s%s%s | p = %f | t = %d ms\n", __func__,
                             "\033[1m", allowed_commands[probs_id[0].second].c_str(), "\033[0m", probs_id[0].first,
-                            (int) std::chrono::duration_cast<std::chrono::milliseconds>(std::chrono::high_resolution_clock::now() - t_start).count());
+                            (int) std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count());
                     fprintf(stdout, "\n");
                 }
 
-                const auto t_end = std::chrono::high_resolution_clock::now();
-
                 audio.clear();
             }
         }

examples/main/main.cpp

@@ -75,7 +75,7 @@ struct whisper_params {
     bool no_timestamps  = false;
 
     std::string language = "en";
-    std::string prompt   = "";
+    std::string prompt;
     std::string model    = "models/ggml-base.en.bin";
 
     std::vector<std::string> fname_inp = {};
@@ -118,7 +118,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-l"    || arg == "--language")       { params.language       = argv[++i]; }
         else if (                  arg == "--prompt")         { params.prompt         = argv[++i]; }
         else if (arg == "-m"    || arg == "--model")          { params.model          = argv[++i]; }
-        else if (arg == "-f"    || arg == "--file")           { params.fname_inp.push_back(argv[++i]); }
+        else if (arg == "-f"    || arg == "--file")           { params.fname_inp.emplace_back(argv[++i]); }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -129,7 +129,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     return true;
 }
 
-void whisper_print_usage(int argc, char ** argv, const whisper_params & params) {
+void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
     fprintf(stderr, "\n");
     fprintf(stderr, "usage: %s [options] file0.wav file1.wav ...\n", argv[0]);
     fprintf(stderr, "\n");
@@ -206,7 +206,7 @@ void whisper_print_segment_callback(struct whisper_context * ctx, int n_new, voi
             const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
             const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
 
-            std::string speaker = "";
+            std::string speaker;
 
             if (params.diarize && pcmf32s.size() == 2) {
                 const int64_t n_samples = pcmf32s[0].size();
@@ -328,7 +328,7 @@ bool output_srt(struct whisper_context * ctx, const char * fname, const whisper_
 // karaoke video generation
 // outputs a bash script that uses ffmpeg to generate a video with the subtitles
 // TODO: font parameter adjustments
-bool output_wts(struct whisper_context * ctx, const char * fname, const char * fname_inp, const whisper_params & params, float t_sec) {
+bool output_wts(struct whisper_context * ctx, const char * fname, const char * fname_inp, const whisper_params & /*params*/, float t_sec) {
     std::ofstream fout(fname);
 
     fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
@@ -377,7 +377,6 @@ bool output_wts(struct whisper_context * ctx, const char * fname, const char * f
             txt_ul = "\\ \\ ";
 
             {
-                int ncnt = 0;
                 for (int k = 0; k < n; ++k) {
                     const auto & token2 = tokens[k];
 
@@ -401,8 +400,6 @@ bool output_wts(struct whisper_context * ctx, const char * fname, const char * f
                             txt_ul += "\\ ";
                         }
                     }
-
-                    ncnt += txt.size();
                 }
 
                 ::replace_all(txt_bg, "'", "\u2019");
@@ -471,7 +468,7 @@ int main(int argc, char ** argv) {
     // initial prompt
     std::vector<whisper_token> prompt_tokens;
 
-    if (params.prompt.size() > 0) {
+    if (!params.prompt.empty()) {
         prompt_tokens.resize(1024);
         prompt_tokens.resize(whisper_tokenize(ctx, params.prompt.c_str(), prompt_tokens.data(), prompt_tokens.size()));
 
@@ -508,14 +505,14 @@ int main(int argc, char ** argv) {
                     }
                 }
 
-                if (drwav_init_memory(&wav, wav_data.data(), wav_data.size(), NULL) == false) {
+                if (drwav_init_memory(&wav, wav_data.data(), wav_data.size(), nullptr) == false) {
                     fprintf(stderr, "error: failed to open WAV file from stdin\n");
                     return 4;
                 }
 
                 fprintf(stderr, "%s: read %zu bytes from stdin\n", __func__, wav_data.size());
             }
-            else if (drwav_init_file(&wav, fname_inp.c_str(), NULL) == false) {
+            else if (drwav_init_file(&wav, fname_inp.c_str(), nullptr) == false) {
                 fprintf(stderr, "error: failed to open '%s' as WAV file\n", fname_inp.c_str());
                 return 5;
             }
@@ -550,11 +547,11 @@ int main(int argc, char ** argv) {
             // convert to mono, float
             pcmf32.resize(n);
             if (wav.channels == 1) {
-                for (int i = 0; i < n; i++) {
+                for (uint64_t i = 0; i < n; i++) {
                     pcmf32[i] = float(pcm16[i])/32768.0f;
                 }
             } else {
-                for (int i = 0; i < n; i++) {
+                for (uint64_t i = 0; i < n; i++) {
                     pcmf32[i] = float(pcm16[2*i] + pcm16[2*i + 1])/65536.0f;
                 }
             }
@@ -565,7 +562,7 @@ int main(int argc, char ** argv) {
 
                 pcmf32s[0].resize(n);
                 pcmf32s[1].resize(n);
-                for (int i = 0; i < n; i++) {
+                for (uint64_t i = 0; i < n; i++) {
                     pcmf32s[0][i] = float(pcm16[2*i])/32768.0f;
                     pcmf32s[1][i] = float(pcm16[2*i + 1])/32768.0f;
                 }
@@ -620,8 +617,8 @@ int main(int argc, char ** argv) {
 
             wparams.speed_up         = params.speed_up;
 
-            wparams.prompt_tokens    = prompt_tokens.size() == 0 ? nullptr : prompt_tokens.data();
-            wparams.prompt_n_tokens  = prompt_tokens.size() == 0 ? 0       : prompt_tokens.size();
+            wparams.prompt_tokens    = prompt_tokens.empty() ? nullptr : prompt_tokens.data();
+            wparams.prompt_n_tokens  = prompt_tokens.empty() ? 0       : prompt_tokens.size();
 
             whisper_print_user_data user_data = { &params, &pcmf32s };
 
@@ -637,7 +634,7 @@ int main(int argc, char ** argv) {
             {
                 static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
 
-                wparams.encoder_begin_callback = [](struct whisper_context * ctx, void * user_data) {
+                wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, void * user_data) {
                     bool is_aborted = *(bool*)user_data;
                     return !is_aborted;
                 };

examples/stream/stream.cpp

@@ -51,7 +51,7 @@ struct whisper_params {
 
     std::string language  = "en";
     std::string model     = "models/ggml-base.en.bin";
-    std::string fname_out = "";
+    std::string fname_out;
 };
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@@ -90,7 +90,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     return true;
 }
 
-void whisper_print_usage(int argc, char ** argv, const whisper_params & params) {
+void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
     fprintf(stderr, "\n");
     fprintf(stderr, "usage: %s [options]\n", argv[0]);
     fprintf(stderr, "\n");
@@ -391,7 +391,7 @@ bool vad_simple(std::vector<float> & pcmf32, int sample_rate, int last_ms, float
     float energy_all  = 0.0f;
     float energy_last = 0.0f;
 
-    for (size_t i = 0; i < n_samples; i++) {
+    for (int i = 0; i < n_samples; i++) {
         energy_all += fabsf(pcmf32[i]);
 
         if (i >= n_samples - n_samples_last) {

examples/talk/gpt-2.cpp

@@ -40,7 +40,7 @@ std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::stri
     // find the longest tokens that form the words:
     std::vector<gpt_vocab::id> tokens;
     for (const auto & word : words) {
-        if (word.size() == 0) continue;
+        if (word.empty()) continue;
 
         int i = 0;
         int n = word.size();
@@ -78,7 +78,7 @@ gpt_vocab::id gpt_sample_top_k_top_p(
         const float * logits,
         int    top_k,
         double top_p,
-        double temp,
+        double /*temp*/,
         std::mt19937 & rng) {
     int n_logits = vocab.id_to_token.size();
 
@@ -86,7 +86,7 @@ gpt_vocab::id gpt_sample_top_k_top_p(
     logits_id.reserve(n_logits);
 
     for (int i = 0; i < n_logits; i++) {
-        logits_id.push_back(std::make_pair(logits[i], i));
+        logits_id.emplace_back(logits[i], i);
     }
 
     // find the top K tokens
@@ -268,7 +268,7 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
             fin.read((char *) &len, sizeof(len));
 
             word.resize(len);
-            fin.read((char *) word.data(), len);
+            fin.read((char *) &word[0], len);
 
             vocab.token_to_id[word] = i;
             vocab.id_to_token[i] = word;
@@ -327,7 +327,7 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
     {
         struct ggml_init_params params;
         params.mem_size   = ctx_size;
-        params.mem_buffer = NULL;
+        params.mem_buffer = nullptr;
 
         model.ctx = ggml_init(params);
         if (!model.ctx) {
@@ -448,7 +448,7 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
             std::string name(length, 0);
             fin.read(&name[0], length);
 
-            if (model.tensors.find(name.data()) == model.tensors.end()) {
+            if (model.tensors.find(name) == model.tensors.end()) {
                 fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
                 return false;
             }
@@ -833,7 +833,7 @@ Me too.
 struct gpt2_context * gpt2_init(const char * path_model) {
     gpt2_context * ctx = new gpt2_context;
 
-    ctx->rng = std::mt19937(time(NULL));
+    ctx->rng = std::mt19937(time(nullptr));
 
     // load the model
     {
@@ -841,6 +841,7 @@ struct gpt2_context * gpt2_init(const char * path_model) {
 
         if (!gpt2_model_load(path_model, ctx->model, ctx->vocab)) {
             fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, path_model);
+            delete ctx;
             return nullptr;
         }
 
@@ -884,9 +885,9 @@ std::string gpt2_gen_text(gpt2_context * ctx, const char * text, int max_tokens)
 
     std::string result;
 
-    for (int i = embd.size(); i < embd_inp.size() + n_predict; i++) {
+    for (int i = embd.size(); i < (int) embd_inp.size() + n_predict; i++) {
         // predict
-        if (embd.size() > 0) {
+        if (!embd.empty()) {
             if (!gpt2_eval(ctx->model, ctx->n_threads, n_past, embd, embd_w, mem_per_token)) {
                 printf("gpt-2: failed to generate text\n");
                 return "";

examples/talk/talk.cpp

@@ -39,7 +39,7 @@ struct whisper_params {
     std::string model_wsp = "models/ggml-base.en.bin";
     std::string model_gpt = "models/ggml-gpt-2-117M.bin";
     std::string speak     = "./examples/talk/speak.sh";
-    std::string fname_out = "";
+    std::string fname_out;
 };
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@@ -79,7 +79,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     return true;
 }
 
-void whisper_print_usage(int argc, char ** argv, const whisper_params & params) {
+void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
     fprintf(stderr, "\n");
     fprintf(stderr, "usage: %s [options]\n", argv[0]);
     fprintf(stderr, "\n");
@@ -397,7 +397,7 @@ bool vad_simple(std::vector<float> & pcmf32, int sample_rate, int last_ms, float
     float energy_all  = 0.0f;
     float energy_last = 0.0f;
 
-    for (size_t i = 0; i < n_samples; i++) {
+    for (int i = 0; i < n_samples; i++) {
         energy_all += fabsf(pcmf32[i]);
 
         if (i >= n_samples - n_samples_last) {
@@ -541,7 +541,6 @@ int main(int argc, char ** argv) {
     bool force_speak = false;
 
     float prob0 = 0.0f;
-    float prob  = 0.0f;
 
     std::vector<float> pcmf32_cur;
     std::vector<float> pcmf32_prompt;
@@ -589,7 +588,7 @@ int main(int argc, char ** argv) {
 
                 audio.get(params.voice_ms, pcmf32_cur);
 
-                std::string text_heard = "";
+                std::string text_heard;
 
                 if (!force_speak) {
                     text_heard = ::trim(::transcribe(ctx_wsp, params, pcmf32_cur, prob0, t_ms));
@@ -611,7 +610,7 @@ int main(int argc, char ** argv) {
                 text_heard = std::regex_replace(text_heard, std::regex("[^a-zA-Z0-9\\.,\\?!\\s\\:\\'\\-]"), "");
 
                 // take first line
-                text_heard = text_heard.substr(0, text_heard.find_first_of("\n"));
+                text_heard = text_heard.substr(0, text_heard.find_first_of('\n'));
 
                 // remove leading and trailing whitespace
                 text_heard = std::regex_replace(text_heard, std::regex("^\\s+"), "");
@@ -641,18 +640,18 @@ int main(int argc, char ** argv) {
 
                     text_to_speak = gpt2_gen_text(ctx_gpt, prompt.c_str(), params.max_tokens);
                     text_to_speak = std::regex_replace(text_to_speak, std::regex("[^a-zA-Z0-9\\.,\\?!\\s\\:\\'\\-]"), "");
-                    text_to_speak = text_to_speak.substr(0, text_to_speak.find_first_of("\n"));
+                    text_to_speak = text_to_speak.substr(0, text_to_speak.find_first_of('\n'));
 
                     // remove first 2 lines of base prompt
                     if (n_iter > 4) {
                         {
-                            const size_t pos = prompt_base.find_first_of("\n");
+                            const size_t pos = prompt_base.find_first_of('\n');
                             if (pos != std::string::npos) {
                                 prompt_base = prompt_base.substr(pos + 1);
                             }
                         }
                         {
-                            const size_t pos = prompt_base.find_first_of("\n");
+                            const size_t pos = prompt_base.find_first_of('\n');
                             if (pos != std::string::npos) {
                                 prompt_base = prompt_base.substr(pos + 1);
                             }

examples/whisper.android/.idea/gradle.xml

@@ -1,5 +1,6 @@
 <?xml version="1.0" encoding="UTF-8"?>
 <project version="4">
+  <component name="GradleMigrationSettings" migrationVersion="1" />
   <component name="GradleSettings">
     <option name="linkedExternalProjectsSettings">
       <GradleProjectSettings>

examples/whisper.android/app/build.gradle

@@ -14,10 +14,6 @@ android {
         versionCode 1
         versionName "1.0"
 
-        ndk {
-            abiFilters 'arm64-v8a', 'x86_64'
-        }
-
         testInstrumentationRunner "androidx.test.runner.AndroidJUnitRunner"
         vectorDrawables {
             useSupportLibrary true
@@ -44,7 +40,7 @@ android {
     composeOptions {
         kotlinCompilerExtensionVersion '1.3.1'
     }
-    ndkVersion "25.0.8528842"
+    ndkVersion "25.1.8937393"
     externalNativeBuild {
         ndkBuild {
             path 'src/main/jni/whisper/Android.mk'

examples/whisper.android/app/src/main/java/com/whispercppdemo/whisper/LibWhisper.kt

@@ -1,8 +1,13 @@
 package com.whispercppdemo.whisper
 
+import android.os.Build
+import android.util.Log
 import kotlinx.coroutines.*
+import java.io.File
 import java.util.concurrent.Executors
 
+private const val LOG_TAG = "LibWhisper"
+
 class WhisperContext private constructor(private var ptr: Long) {
     // Meet Whisper C++ constraint: Don't access from more than one thread at a time.
     private val scope: CoroutineScope = CoroutineScope(
@@ -47,7 +52,27 @@ class WhisperContext private constructor(private var ptr: Long) {
 private class WhisperLib {
     companion object {
         init {
-            System.loadLibrary("whisper")
+            Log.d(LOG_TAG, "Primary ABI: ${Build.SUPPORTED_ABIS[0]}")
+            var loadVfpv4 = false
+            if (isArmEabiV7a()) {
+                // armeabi-v7a needs runtime detection support
+                val cpuInfo = cpuInfo()
+                cpuInfo?.let {
+                    Log.d(LOG_TAG, "CPU info: $cpuInfo")
+                    if (cpuInfo.contains("vfpv4")) {
+                        Log.d(LOG_TAG, "CPU supports vfpv4")
+                        loadVfpv4 = true
+                    }
+                }
+            }
+
+            if (loadVfpv4) {
+                Log.d(LOG_TAG, "Loading libwhisper_vfpv4.so")
+                System.loadLibrary("whisper_vfpv4")
+            } else {
+                Log.d(LOG_TAG, "Loading libwhisper.so")
+                System.loadLibrary("whisper")
+            }
         }
 
         // JNI methods
@@ -59,3 +84,17 @@ private class WhisperLib {
     }
 }
 
+private fun isArmEabiV7a(): Boolean {
+    return Build.SUPPORTED_ABIS[0].equals("armeabi-v7a")
+}
+
+private fun cpuInfo(): String? {
+    return try {
+        File("/proc/cpuinfo").inputStream().bufferedReader().use {
+            it.readText()
+        }
+    } catch (e: Exception) {
+        Log.w(LOG_TAG, "Couldn't read /proc/cpuinfo", e)
+        null
+    }
+}

examples/whisper.android/app/src/main/jni/whisper/Android.mk

@@ -1,22 +1,15 @@
 LOCAL_PATH := $(call my-dir)
 include $(CLEAR_VARS)
-WHISPER_LIB_DIR := $(LOCAL_PATH)/../../../../../../../
-LOCAL_LDLIBS    := -llog
 LOCAL_MODULE    := libwhisper
+include $(LOCAL_PATH)/Whisper.mk
+include $(BUILD_SHARED_LIBRARY)
 
-# Make the final output library smaller by only keeping the symbols referenced from the app.
-ifneq ($(APP_OPTIM),debug)
-    LOCAL_CFLAGS += -fvisibility=hidden -fvisibility-inlines-hidden
-    LOCAL_CFLAGS += -ffunction-sections -fdata-sections
-    LOCAL_LDFLAGS += -Wl,--gc-sections
-    LOCAL_LDFLAGS += -Wl,--exclude-libs,ALL
-    LOCAL_LDFLAGS += -flto
-endif
-
-LOCAL_CFLAGS    += -DSTDC_HEADERS -std=c11 -I $(WHISPER_LIB_DIR)
-LOCAL_CPPFLAGS  += -std=c++11
-LOCAL_SRC_FILES := $(WHISPER_LIB_DIR)/ggml.c \
-                   $(WHISPER_LIB_DIR)/whisper.cpp \
-                   $(LOCAL_PATH)/jni.c
-
-include $(BUILD_SHARED_LIBRARY)
+ifeq ($(TARGET_ARCH_ABI),armeabi-v7a)
+	include $(CLEAR_VARS)
+	LOCAL_MODULE    := libwhisper_vfpv4
+	include $(LOCAL_PATH)/Whisper.mk
+	# Allow building NEON FMA code.
+	# https://android.googlesource.com/platform/ndk/+/master/sources/android/cpufeatures/cpu-features.h
+	LOCAL_CFLAGS += -mfpu=neon-vfpv4
+	include $(BUILD_SHARED_LIBRARY)
+endif

examples/whisper.android/app/src/main/jni/whisper/Whisper.mk

@@ -0,0 +1,18 @@
+WHISPER_LIB_DIR := $(LOCAL_PATH)/../../../../../../../
+LOCAL_LDLIBS    := -llog
+
+# Make the final output library smaller by only keeping the symbols referenced from the app.
+ifneq ($(APP_OPTIM),debug)
+    LOCAL_CFLAGS += -O3
+    LOCAL_CFLAGS += -fvisibility=hidden -fvisibility-inlines-hidden
+    LOCAL_CFLAGS += -ffunction-sections -fdata-sections
+    LOCAL_LDFLAGS += -Wl,--gc-sections
+    LOCAL_LDFLAGS += -Wl,--exclude-libs,ALL
+    LOCAL_LDFLAGS += -flto
+endif
+
+LOCAL_CFLAGS    += -DSTDC_HEADERS -std=c11 -I $(WHISPER_LIB_DIR)
+LOCAL_CPPFLAGS  += -std=c++11
+LOCAL_SRC_FILES := $(WHISPER_LIB_DIR)/ggml.c \
+                   $(WHISPER_LIB_DIR)/whisper.cpp \
+                   $(LOCAL_PATH)/jni.c

examples/whisper.objc/README.md

@@ -19,3 +19,8 @@ open whisper.cpp/examples/whisper.objc/whisper.objc.xcodeproj/
 Make sure to build the project in `Release`:
 
 <img width="947" alt="image" src="https://user-images.githubusercontent.com/1991296/197382607-9e1e6d1b-79fa-496f-9d16-b71dc1535701.png">
+
+Also, don't forget to add the `-DGGML_USE_ACCELERATE` compiler flag in Build Phases.
+This can significantly improve the performance of the transcription:
+
+<img width="1072" alt="image" src="https://user-images.githubusercontent.com/1991296/208511239-8d7cdbd1-aa48-41b5-becd-ca288d53cc07.png">

examples/whisper.swiftui/README.md

@@ -0,0 +1,12 @@
+A sample SwiftUI app using [whisper.cpp](https://github.com/ggerganov/whisper.cpp/) to do voice-to-text transcriptions.
+See also: [whisper.objc](https://github.com/ggerganov/whisper.cpp/tree/master/examples/whisper.objc).
+
+To use:
+
+1. Select a model from the [whisper.cpp repository](https://github.com/ggerganov/whisper.cpp/tree/master/models).[^1]
+2. Add the model to "whisper.swiftui.demo/Resources/models" via Xcode.
+3. Select a sample audio file (for example, [jfk.wav](https://github.com/ggerganov/whisper.cpp/raw/master/samples/jfk.wav)).
+4. Add the model to "whisper.swiftui.demo/Resources/samples" via Xcode.
+5. Select the "release" build configuration under "Run", then deploy and run to your device.
+
+[^1]: I recommend the tiny, base or small models for running on an iOS device.

examples/whisper.swiftui/whisper.cpp.swift/LibWhisper.swift

@@ -0,0 +1,70 @@
+import Foundation
+
+enum WhisperError: Error {
+    case couldNotInitializeContext
+}
+
+// Meet Whisper C++ constraint: Don't access from more than one thread at a time.
+actor WhisperContext {
+    private var context: OpaquePointer
+    
+    init(context: OpaquePointer) {
+        self.context = context
+    }
+    
+    deinit {
+        whisper_free(context)
+    }
+    
+    func fullTranscribe(samples: [Float]) {
+        // Leave 2 processors free (i.e. the high-efficiency cores).
+        let maxThreads = max(1, min(8, cpuCount() - 2))
+        print("Selecting \(maxThreads) threads")
+        var params = whisper_full_default_params(WHISPER_SAMPLING_GREEDY)
+        "en".withCString { en in
+            // Adapted from whisper.objc
+            params.print_realtime = true
+            params.print_progress = false
+            params.print_timestamps = true
+            params.print_special = false
+            params.translate = false
+            params.language = en
+            params.n_threads = Int32(maxThreads)
+            params.offset_ms = 0
+            params.no_context = true
+            params.single_segment = false
+            
+            whisper_reset_timings(context)
+            print("About to run whisper_full")
+            samples.withUnsafeBufferPointer { samples in
+                if (whisper_full(context, params, samples.baseAddress, Int32(samples.count)) != 0) {
+                    print("Failed to run the model")
+                } else {
+                    whisper_print_timings(context)
+                }
+            }
+        }
+    }
+    
+    func getTranscription() -> String {
+        var transcription = ""
+        for i in 0..<whisper_full_n_segments(context) {
+            transcription += String.init(cString: whisper_full_get_segment_text(context, i))
+        }
+        return transcription
+    }
+    
+    static func createContext(path: String) throws -> WhisperContext {
+        let context = whisper_init(path)
+        if let context {
+            return WhisperContext(context: context)
+        } else {
+            print("Couldn't load model at \(path)")
+            throw WhisperError.couldNotInitializeContext
+        }
+    }
+}
+
+fileprivate func cpuCount() -> Int {
+    ProcessInfo.processInfo.processorCount
+}

examples/whisper.swiftui/whisper.cpp.swift/WhisperCppDemo-Bridging-Header.h

@@ -0,0 +1,4 @@
+//
+//  Use this file to import your target's public headers that you would like to expose to Swift.
+//
+#import "whisper.h"

examples/whisper.swiftui/whisper.swiftui.demo/Models/WhisperState.swift

@@ -0,0 +1,162 @@
+import Foundation
+import SwiftUI
+import AVFoundation
+
+@MainActor
+class WhisperState: NSObject, ObservableObject, AVAudioRecorderDelegate {
+    @Published var isModelLoaded = false
+    @Published var messageLog = ""
+    @Published var canTranscribe = false
+    @Published var isRecording = false
+    
+    private var whisperContext: WhisperContext?
+    private let recorder = Recorder()
+    private var recordedFile: URL? = nil
+    private var audioPlayer: AVAudioPlayer?
+    
+    private var modelUrl: URL? {
+        Bundle.main.url(forResource: "ggml-tiny.en", withExtension: "bin", subdirectory: "models")
+    }
+    
+    private var sampleUrl: URL? {
+        Bundle.main.url(forResource: "jfk", withExtension: "wav", subdirectory: "samples")
+    }
+    
+    private enum LoadError: Error {
+        case couldNotLocateModel
+    }
+    
+    override init() {
+        super.init()
+        do {
+            try loadModel()
+            canTranscribe = true
+        } catch {
+            print(error.localizedDescription)
+            messageLog += "\(error.localizedDescription)\n"
+        }
+    }
+    
+    private func loadModel() throws {
+        messageLog += "Loading model...\n"
+        if let modelUrl {
+            whisperContext = try WhisperContext.createContext(path: modelUrl.path())
+            messageLog += "Loaded model \(modelUrl.lastPathComponent)\n"
+        } else {
+            messageLog += "Could not locate model\n"
+        }
+    }
+    
+    func transcribeSample() async {
+        if let sampleUrl {
+            await transcribeAudio(sampleUrl)
+        } else {
+            messageLog += "Could not locate sample\n"
+        }
+    }
+    
+    private func transcribeAudio(_ url: URL) async {
+        if (!canTranscribe) {
+            return
+        }
+        guard let whisperContext else {
+            return
+        }
+        
+        do {
+            canTranscribe = false
+            messageLog += "Reading wave samples...\n"
+            let data = try readAudioSamples(url)
+            messageLog += "Transcribing data...\n"
+            await whisperContext.fullTranscribe(samples: data)
+            let text = await whisperContext.getTranscription()
+            messageLog += "Done: \(text)\n"
+        } catch {
+            print(error.localizedDescription)
+            messageLog += "\(error.localizedDescription)\n"
+        }
+        
+        canTranscribe = true
+    }
+    
+    private func readAudioSamples(_ url: URL) throws -> [Float] {
+        stopPlayback()
+        try startPlayback(url)
+        return try decodeWaveFile(url)
+    }
+    
+    func toggleRecord() async {
+        if isRecording {
+            await recorder.stopRecording()
+            isRecording = false
+            if let recordedFile {
+                await transcribeAudio(recordedFile)
+            }
+        } else {
+            requestRecordPermission { granted in
+                if granted {
+                    Task {
+                        do {
+                            self.stopPlayback()
+                            let file = try FileManager.default.url(for: .documentDirectory, in: .userDomainMask, appropriateFor: nil, create: true)
+                                .appending(path: "output.wav")
+                            try await self.recorder.startRecording(toOutputFile: file, delegate: self)
+                            self.isRecording = true
+                            self.recordedFile = file
+                        } catch {
+                            print(error.localizedDescription)
+                            self.messageLog += "\(error.localizedDescription)\n"
+                            self.isRecording = false
+                        }
+                    }
+                }
+            }
+        }
+    }
+    
+    private func requestRecordPermission(response: @escaping (Bool) -> Void) {
+#if os(macOS)
+        response(true)
+#else
+        AVAudioSession.sharedInstance().requestRecordPermission { granted in
+            response(granted)
+        }
+#endif
+    }
+    
+    private func startPlayback(_ url: URL) throws {
+        audioPlayer = try AVAudioPlayer(contentsOf: url)
+        audioPlayer?.play()
+    }
+    
+    private func stopPlayback() {
+        audioPlayer?.stop()
+        audioPlayer = nil
+    }
+    
+    // MARK: AVAudioRecorderDelegate
+    
+    nonisolated func audioRecorderEncodeErrorDidOccur(_ recorder: AVAudioRecorder, error: Error?) {
+        if let error {
+            Task {
+                await handleRecError(error)
+            }
+        }
+    }
+    
+    private func handleRecError(_ error: Error) {
+        print(error.localizedDescription)
+        messageLog += "\(error.localizedDescription)\n"
+        isRecording = false
+    }
+    
+    nonisolated func audioRecorderDidFinishRecording(_ recorder: AVAudioRecorder, successfully flag: Bool) {
+        Task {
+            await onDidFinishRecording()
+        }
+    }
+    
+    private func onDidFinishRecording() {
+        isRecording = false
+    }
+}

examples/whisper.swiftui/whisper.swiftui.demo/Supporting files/Assets.xcassets/AccentColor.colorset/Contents.json

@@ -0,0 +1,11 @@
+{
+  "colors" : [
+    {
+      "idiom" : "universal"
+    }
+  ],
+  "info" : {
+    "author" : "xcode",
+    "version" : 1
+  }
+}

examples/whisper.swiftui/whisper.swiftui.demo/Supporting files/Assets.xcassets/AppIcon.appiconset/Contents.json

@@ -0,0 +1,63 @@
+{
+  "images" : [
+    {
+      "idiom" : "universal",
+      "platform" : "ios",
+      "size" : "1024x1024"
+    },
+    {
+      "idiom" : "mac",
+      "scale" : "1x",
+      "size" : "16x16"
+    },
+    {
+      "idiom" : "mac",
+      "scale" : "2x",
+      "size" : "16x16"
+    },
+    {
+      "idiom" : "mac",
+      "scale" : "1x",
+      "size" : "32x32"
+    },
+    {
+      "idiom" : "mac",
+      "scale" : "2x",
+      "size" : "32x32"
+    },
+    {
+      "idiom" : "mac",
+      "scale" : "1x",
+      "size" : "128x128"
+    },
+    {
+      "idiom" : "mac",
+      "scale" : "2x",
+      "size" : "128x128"
+    },
+    {
+      "idiom" : "mac",
+      "scale" : "1x",
+      "size" : "256x256"
+    },
+    {
+      "idiom" : "mac",
+      "scale" : "2x",
+      "size" : "256x256"
+    },
+    {
+      "idiom" : "mac",
+      "scale" : "1x",
+      "size" : "512x512"
+    },
+    {
+      "idiom" : "mac",
+      "scale" : "2x",
+      "size" : "512x512"
+    }
+  ],
+  "info" : {
+    "author" : "xcode",
+    "version" : 1
+  }
+}

examples/whisper.swiftui/whisper.swiftui.demo/Supporting files/Assets.xcassets/Contents.json

@@ -0,0 +1,6 @@
+{
+  "info" : {
+    "author" : "xcode",
+    "version" : 1
+  }
+}

examples/whisper.swiftui/whisper.swiftui.demo/Supporting files/Preview Content/Preview Assets.xcassets/Contents.json

@@ -0,0 +1,6 @@
+{
+  "info" : {
+    "author" : "xcode",
+    "version" : 1
+  }
+}

examples/whisper.swiftui/whisper.swiftui.demo/Supporting files/WhisperCppDemo.entitlements

@@ -0,0 +1,12 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+	<key>com.apple.security.app-sandbox</key>
+	<true/>
+	<key>com.apple.security.device.audio-input</key>
+	<true/>
+	<key>com.apple.security.files.user-selected.read-only</key>
+	<true/>
+</dict>
+</plist>

examples/whisper.swiftui/whisper.swiftui.demo/UI/ContentView.swift

@@ -0,0 +1,43 @@
+import SwiftUI
+import AVFoundation
+
+struct ContentView: View {
+    @StateObject var whisperState = WhisperState()
+    
+    var body: some View {
+        NavigationStack {
+            VStack {
+                HStack {
+                    Button("Transcribe", action: {
+                        Task {
+                            await whisperState.transcribeSample()
+                        }
+                    })
+                    .buttonStyle(.bordered)
+                    .disabled(!whisperState.canTranscribe)
+                    
+                    Button(whisperState.isRecording ? "Stop recording" : "Start recording", action: {
+                        Task {
+                            await whisperState.toggleRecord()
+                        }
+                    })
+                    .buttonStyle(.bordered)
+                    .disabled(!whisperState.canTranscribe)
+                }
+                
+                ScrollView {
+                    Text(verbatim: whisperState.messageLog)
+                        .frame(maxWidth: .infinity, alignment: .leading)
+                }
+            }
+            .navigationTitle("Whisper SwiftUI Demo")
+            .padding()
+        }
+    }
+}
+
+struct ContentView_Previews: PreviewProvider {
+    static var previews: some View {
+        ContentView()
+    }
+}

examples/whisper.swiftui/whisper.swiftui.demo/Utils/Recorder.swift

@@ -0,0 +1,35 @@
+import Foundation
+import AVFoundation
+
+actor Recorder {
+    private var recorder: AVAudioRecorder?
+    
+    enum RecorderError: Error {
+        case couldNotStartRecording
+    }
+    
+    func startRecording(toOutputFile url: URL, delegate: AVAudioRecorderDelegate?) throws {
+        let recordSettings: [String : Any] = [
+            AVFormatIDKey: Int(kAudioFormatLinearPCM),
+            AVSampleRateKey: 16000.0,
+            AVNumberOfChannelsKey: 1,
+            AVEncoderAudioQualityKey: AVAudioQuality.high.rawValue
+        ]
+#if !os(macOS)
+        let session = AVAudioSession.sharedInstance()
+        try session.setCategory(.playAndRecord, mode: .default)
+#endif
+        let recorder = try AVAudioRecorder(url: url, settings: recordSettings)
+        recorder.delegate = delegate
+        if recorder.record() == false {
+            print("Could not start recording")
+            throw RecorderError.couldNotStartRecording
+        }
+        self.recorder = recorder
+    }
+    
+    func stopRecording() {
+        recorder?.stop()
+        recorder = nil
+    }
+}

examples/whisper.swiftui/whisper.swiftui.demo/Utils/RiffWaveUtils.swift

@@ -0,0 +1,12 @@
+import Foundation
+
+func decodeWaveFile(_ url: URL) throws -> [Float] {
+    let data = try Data(contentsOf: url)
+    let floats = stride(from: 44, to: data.count, by: 2).map {
+        return data[$0..<$0 + 2].withUnsafeBytes {
+            let short = Int16(littleEndian: $0.load(as: Int16.self))
+            return max(-1.0, min(Float(short) / 32767.0, 1.0))
+        }
+    }
+    return floats
+}

examples/whisper.swiftui/whisper.swiftui.demo/WhisperCppDemoApp.swift

@@ -0,0 +1,10 @@
+import SwiftUI
+
+@main
+struct WhisperCppDemoApp: App {
+    var body: some Scene {
+        WindowGroup {
+            ContentView()
+        }
+    }
+}

examples/whisper.swiftui/whisper.swiftui.xcodeproj/.gitignore

@@ -0,0 +1 @@
+xcuserdata/

examples/whisper.swiftui/whisper.swiftui.xcodeproj/project.pbxproj

@@ -0,0 +1,467 @@
+// !$*UTF8*$!
+{
+	archiveVersion = 1;
+	classes = {
+	};
+	objectVersion = 56;
+	objects = {
+
+/* Begin PBXBuildFile section */
+		0A8E49002954B3F100704C1B /* README.md in Resources */ = {isa = PBXBuildFile; fileRef = 0A8E48FF2954B3F100704C1B /* README.md */; };
+		0AA751482953AC2E001EE061 /* samples in Resources */ = {isa = PBXBuildFile; fileRef = 0AA751462953AC2E001EE061 /* samples */; };
+		0AA751492953AC2E001EE061 /* models in Resources */ = {isa = PBXBuildFile; fileRef = 0AA751472953AC2E001EE061 /* models */; };
+		0AA7514C2953B569001EE061 /* RiffWaveUtils.swift in Sources */ = {isa = PBXBuildFile; fileRef = 0AA7514B2953B569001EE061 /* RiffWaveUtils.swift */; };
+		0AA7514E2953D958001EE061 /* Recorder.swift in Sources */ = {isa = PBXBuildFile; fileRef = 0AA7514D2953D958001EE061 /* Recorder.swift */; };
+		0AAC5D9B29539CCF003032C3 /* WhisperCppDemoApp.swift in Sources */ = {isa = PBXBuildFile; fileRef = 0AAC5D9A29539CCF003032C3 /* WhisperCppDemoApp.swift */; };
+		0AAC5D9D29539CCF003032C3 /* ContentView.swift in Sources */ = {isa = PBXBuildFile; fileRef = 0AAC5D9C29539CCF003032C3 /* ContentView.swift */; };
+		0AAC5D9F29539CD0003032C3 /* Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 0AAC5D9E29539CD0003032C3 /* Assets.xcassets */; };
+		0AAC5DA329539CD0003032C3 /* Preview Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 0AAC5DA229539CD0003032C3 /* Preview Assets.xcassets */; };
+		0AAC5DCB29539EB1003032C3 /* whisper.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 0AAC5DC729539EB0003032C3 /* whisper.cpp */; settings = {COMPILER_FLAGS = "-Wno-shorten-64-to-32"; }; };
+		0AAC5DCC29539EB1003032C3 /* ggml.c in Sources */ = {isa = PBXBuildFile; fileRef = 0AAC5DC929539EB0003032C3 /* ggml.c */; settings = {COMPILER_FLAGS = "-DGGML_USE_ACCELERATE -Wno-shorten-64-to-32"; }; };
+		0AAC5DCE2953A05C003032C3 /* WhisperState.swift in Sources */ = {isa = PBXBuildFile; fileRef = 0AAC5DCD2953A05C003032C3 /* WhisperState.swift */; };
+		0AAC5DD12953A394003032C3 /* LibWhisper.swift in Sources */ = {isa = PBXBuildFile; fileRef = 0AAC5DD02953A394003032C3 /* LibWhisper.swift */; };
+/* End PBXBuildFile section */
+
+/* Begin PBXFileReference section */
+		0A8E48FF2954B3F100704C1B /* README.md */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = net.daringfireball.markdown; path = README.md; sourceTree = "<group>"; };
+		0AA751462953AC2E001EE061 /* samples */ = {isa = PBXFileReference; lastKnownFileType = folder; path = samples; sourceTree = "<group>"; };
+		0AA751472953AC2E001EE061 /* models */ = {isa = PBXFileReference; lastKnownFileType = folder; path = models; sourceTree = "<group>"; };
+		0AA7514B2953B569001EE061 /* RiffWaveUtils.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = RiffWaveUtils.swift; sourceTree = "<group>"; };
+		0AA7514D2953D958001EE061 /* Recorder.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = Recorder.swift; sourceTree = "<group>"; };
+		0AAC5D9729539CCF003032C3 /* whisper.swiftui.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = whisper.swiftui.app; sourceTree = BUILT_PRODUCTS_DIR; };
+		0AAC5D9A29539CCF003032C3 /* WhisperCppDemoApp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = WhisperCppDemoApp.swift; sourceTree = "<group>"; };
+		0AAC5D9C29539CCF003032C3 /* ContentView.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ContentView.swift; sourceTree = "<group>"; };
+		0AAC5D9E29539CD0003032C3 /* Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = Assets.xcassets; sourceTree = "<group>"; };
+		0AAC5DA029539CD0003032C3 /* WhisperCppDemo.entitlements */ = {isa = PBXFileReference; lastKnownFileType = text.plist.entitlements; path = WhisperCppDemo.entitlements; sourceTree = "<group>"; };
+		0AAC5DA229539CD0003032C3 /* Preview Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = "Preview Assets.xcassets"; sourceTree = "<group>"; };
+		0AAC5DC629539EAF003032C3 /* WhisperCppDemo-Bridging-Header.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = "WhisperCppDemo-Bridging-Header.h"; sourceTree = "<group>"; };
+		0AAC5DC729539EB0003032C3 /* whisper.cpp */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.cpp; name = whisper.cpp; path = ../../../whisper.cpp; sourceTree = "<group>"; };
+		0AAC5DC829539EB0003032C3 /* whisper.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = whisper.h; path = ../../../whisper.h; sourceTree = "<group>"; };
+		0AAC5DC929539EB0003032C3 /* ggml.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = ggml.c; path = ../../../ggml.c; sourceTree = "<group>"; };
+		0AAC5DCA29539EB0003032C3 /* ggml.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = ggml.h; path = ../../../ggml.h; sourceTree = "<group>"; };
+		0AAC5DCD2953A05C003032C3 /* WhisperState.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = WhisperState.swift; sourceTree = "<group>"; };
+		0AAC5DD02953A394003032C3 /* LibWhisper.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LibWhisper.swift; sourceTree = "<group>"; };
+/* End PBXFileReference section */
+
+/* Begin PBXFrameworksBuildPhase section */
+		0AAC5D9429539CCF003032C3 /* Frameworks */ = {
+			isa = PBXFrameworksBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+/* End PBXFrameworksBuildPhase section */
+
+/* Begin PBXGroup section */
+		0AA7513F2953AB32001EE061 /* Models */ = {
+			isa = PBXGroup;
+			children = (
+				0AAC5DCD2953A05C003032C3 /* WhisperState.swift */,
+			);
+			path = Models;
+			sourceTree = "<group>";
+		};
+		0AA751402953ABA6001EE061 /* Resources */ = {
+			isa = PBXGroup;
+			children = (
+				0AA751472953AC2E001EE061 /* models */,
+				0AA751462953AC2E001EE061 /* samples */,
+			);
+			path = Resources;
+			sourceTree = "<group>";
+		};
+		0AA7514A2953B561001EE061 /* Utils */ = {
+			isa = PBXGroup;
+			children = (
+				0AA7514B2953B569001EE061 /* RiffWaveUtils.swift */,
+				0AA7514D2953D958001EE061 /* Recorder.swift */,
+			);
+			path = Utils;
+			sourceTree = "<group>";
+		};
+		0AAC5D8E29539CCF003032C3 = {
+			isa = PBXGroup;
+			children = (
+				0A8E48FF2954B3F100704C1B /* README.md */,
+				0AAC5DC529539E89003032C3 /* whisper.cpp */,
+				0AAC5DCF2953A36C003032C3 /* whisper.cpp.swift */,
+				0AAC5D9929539CCF003032C3 /* whisper.swiftui.demo */,
+				0AAC5D9829539CCF003032C3 /* Products */,
+			);
+			sourceTree = "<group>";
+		};
+		0AAC5D9829539CCF003032C3 /* Products */ = {
+			isa = PBXGroup;
+			children = (
+				0AAC5D9729539CCF003032C3 /* whisper.swiftui.app */,
+			);
+			name = Products;
+			sourceTree = "<group>";
+		};
+		0AAC5D9929539CCF003032C3 /* whisper.swiftui.demo */ = {
+			isa = PBXGroup;
+			children = (
+				0AA7514A2953B561001EE061 /* Utils */,
+				0AA751402953ABA6001EE061 /* Resources */,
+				0AA7513F2953AB32001EE061 /* Models */,
+				0AAC5DD32953A9ED003032C3 /* Supporting files */,
+				0AAC5DD22953A9E3003032C3 /* UI */,
+				0AAC5D9A29539CCF003032C3 /* WhisperCppDemoApp.swift */,
+			);
+			path = whisper.swiftui.demo;
+			sourceTree = "<group>";
+		};
+		0AAC5DA129539CD0003032C3 /* Preview Content */ = {
+			isa = PBXGroup;
+			children = (
+				0AAC5DA229539CD0003032C3 /* Preview Assets.xcassets */,
+			);
+			name = "Preview Content";
+			path = "../Preview Content";
+			sourceTree = "<group>";
+		};
+		0AAC5DC529539E89003032C3 /* whisper.cpp */ = {
+			isa = PBXGroup;
+			children = (
+				0AAC5DC929539EB0003032C3 /* ggml.c */,
+				0AAC5DCA29539EB0003032C3 /* ggml.h */,
+				0AAC5DC729539EB0003032C3 /* whisper.cpp */,
+				0AAC5DC829539EB0003032C3 /* whisper.h */,
+			);
+			path = whisper.cpp;
+			sourceTree = "<group>";
+		};
+		0AAC5DCF2953A36C003032C3 /* whisper.cpp.swift */ = {
+			isa = PBXGroup;
+			children = (
+				0AAC5DC629539EAF003032C3 /* WhisperCppDemo-Bridging-Header.h */,
+				0AAC5DD02953A394003032C3 /* LibWhisper.swift */,
+			);
+			path = whisper.cpp.swift;
+			sourceTree = "<group>";
+		};
+		0AAC5DD22953A9E3003032C3 /* UI */ = {
+			isa = PBXGroup;
+			children = (
+				0AAC5D9C29539CCF003032C3 /* ContentView.swift */,
+			);
+			path = UI;
+			sourceTree = "<group>";
+		};
+		0AAC5DD32953A9ED003032C3 /* Supporting files */ = {
+			isa = PBXGroup;
+			children = (
+				0AAC5D9E29539CD0003032C3 /* Assets.xcassets */,
+				0AAC5DA029539CD0003032C3 /* WhisperCppDemo.entitlements */,
+				0AAC5DA129539CD0003032C3 /* Preview Content */,
+			);
+			path = "Supporting files";
+			sourceTree = "<group>";
+		};
+/* End PBXGroup section */
+
+/* Begin PBXNativeTarget section */
+		0AAC5D9629539CCF003032C3 /* whisper.swiftui */ = {
+			isa = PBXNativeTarget;
+			buildConfigurationList = 0AAC5DBC29539CD0003032C3 /* Build configuration list for PBXNativeTarget "whisper.swiftui" */;
+			buildPhases = (
+				0AAC5D9329539CCF003032C3 /* Sources */,
+				0AAC5D9429539CCF003032C3 /* Frameworks */,
+				0AAC5D9529539CCF003032C3 /* Resources */,
+			);
+			buildRules = (
+			);
+			dependencies = (
+			);
+			name = whisper.swiftui;
+			productName = WhisperCppDemo;
+			productReference = 0AAC5D9729539CCF003032C3 /* whisper.swiftui.app */;
+			productType = "com.apple.product-type.application";
+		};
+/* End PBXNativeTarget section */
+
+/* Begin PBXProject section */
+		0AAC5D8F29539CCF003032C3 /* Project object */ = {
+			isa = PBXProject;
+			attributes = {
+				BuildIndependentTargetsInParallel = 1;
+				LastSwiftUpdateCheck = 1410;
+				LastUpgradeCheck = 1410;
+				TargetAttributes = {
+					0AAC5D9629539CCF003032C3 = {
+						CreatedOnToolsVersion = 14.1;
+						LastSwiftMigration = 1410;
+					};
+				};
+			};
+			buildConfigurationList = 0AAC5D9229539CCF003032C3 /* Build configuration list for PBXProject "whisper.swiftui" */;
+			compatibilityVersion = "Xcode 14.0";
+			developmentRegion = en;
+			hasScannedForEncodings = 0;
+			knownRegions = (
+				en,
+				Base,
+			);
+			mainGroup = 0AAC5D8E29539CCF003032C3;
+			productRefGroup = 0AAC5D9829539CCF003032C3 /* Products */;
+			projectDirPath = "";
+			projectRoot = "";
+			targets = (
+				0AAC5D9629539CCF003032C3 /* whisper.swiftui */,
+			);
+		};
+/* End PBXProject section */
+
+/* Begin PBXResourcesBuildPhase section */
+		0AAC5D9529539CCF003032C3 /* Resources */ = {
+			isa = PBXResourcesBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+				0AA751482953AC2E001EE061 /* samples in Resources */,
+				0AAC5DA329539CD0003032C3 /* Preview Assets.xcassets in Resources */,
+				0A8E49002954B3F100704C1B /* README.md in Resources */,
+				0AA751492953AC2E001EE061 /* models in Resources */,
+				0AAC5D9F29539CD0003032C3 /* Assets.xcassets in Resources */,
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+/* End PBXResourcesBuildPhase section */
+
+/* Begin PBXSourcesBuildPhase section */
+		0AAC5D9329539CCF003032C3 /* Sources */ = {
+			isa = PBXSourcesBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+				0AAC5D9D29539CCF003032C3 /* ContentView.swift in Sources */,
+				0AAC5D9B29539CCF003032C3 /* WhisperCppDemoApp.swift in Sources */,
+				0AAC5DCC29539EB1003032C3 /* ggml.c in Sources */,
+				0AAC5DCE2953A05C003032C3 /* WhisperState.swift in Sources */,
+				0AAC5DD12953A394003032C3 /* LibWhisper.swift in Sources */,
+				0AA7514C2953B569001EE061 /* RiffWaveUtils.swift in Sources */,
+				0AAC5DCB29539EB1003032C3 /* whisper.cpp in Sources */,
+				0AA7514E2953D958001EE061 /* Recorder.swift in Sources */,
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+/* End PBXSourcesBuildPhase section */
+
+/* Begin XCBuildConfiguration section */
+		0AAC5DBA29539CD0003032C3 /* Debug */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				ALWAYS_SEARCH_USER_PATHS = NO;
+				CLANG_ANALYZER_NONNULL = YES;
+				CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE;
+				CLANG_CXX_LANGUAGE_STANDARD = "gnu++20";
+				CLANG_ENABLE_MODULES = YES;
+				CLANG_ENABLE_OBJC_ARC = YES;
+				CLANG_ENABLE_OBJC_WEAK = YES;
+				CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES;
+				CLANG_WARN_BOOL_CONVERSION = YES;
+				CLANG_WARN_COMMA = YES;
+				CLANG_WARN_CONSTANT_CONVERSION = YES;
+				CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES;
+				CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR;
+				CLANG_WARN_DOCUMENTATION_COMMENTS = YES;
+				CLANG_WARN_EMPTY_BODY = YES;
+				CLANG_WARN_ENUM_CONVERSION = YES;
+				CLANG_WARN_INFINITE_RECURSION = YES;
+				CLANG_WARN_INT_CONVERSION = YES;
+				CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES;
+				CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES;
+				CLANG_WARN_OBJC_LITERAL_CONVERSION = YES;
+				CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR;
+				CLANG_WARN_QUOTED_INCLUDE_IN_FRAMEWORK_HEADER = YES;
+				CLANG_WARN_RANGE_LOOP_ANALYSIS = YES;
+				CLANG_WARN_STRICT_PROTOTYPES = YES;
+				CLANG_WARN_SUSPICIOUS_MOVE = YES;
+				CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE;
+				CLANG_WARN_UNREACHABLE_CODE = YES;
+				CLANG_WARN__DUPLICATE_METHOD_MATCH = YES;
+				COPY_PHASE_STRIP = NO;
+				DEBUG_INFORMATION_FORMAT = dwarf;
+				ENABLE_STRICT_OBJC_MSGSEND = YES;
+				ENABLE_TESTABILITY = YES;
+				GCC_C_LANGUAGE_STANDARD = gnu11;
+				GCC_DYNAMIC_NO_PIC = NO;
+				GCC_NO_COMMON_BLOCKS = YES;
+				GCC_OPTIMIZATION_LEVEL = 0;
+				GCC_PREPROCESSOR_DEFINITIONS = (
+					"DEBUG=1",
+					"$(inherited)",
+				);
+				GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
+				GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR;
+				GCC_WARN_UNDECLARED_SELECTOR = YES;
+				GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
+				GCC_WARN_UNUSED_FUNCTION = YES;
+				GCC_WARN_UNUSED_VARIABLE = YES;
+				INFOPLIST_KEY_NSMicrophoneUsageDescription = "Needed to transcribe audio";
+				MTL_ENABLE_DEBUG_INFO = INCLUDE_SOURCE;
+				MTL_FAST_MATH = YES;
+				ONLY_ACTIVE_ARCH = YES;
+				SWIFT_ACTIVE_COMPILATION_CONDITIONS = DEBUG;
+				SWIFT_OPTIMIZATION_LEVEL = "-Onone";
+			};
+			name = Debug;
+		};
+		0AAC5DBB29539CD0003032C3 /* Release */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				ALWAYS_SEARCH_USER_PATHS = NO;
+				CLANG_ANALYZER_NONNULL = YES;
+				CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE;
+				CLANG_CXX_LANGUAGE_STANDARD = "gnu++20";
+				CLANG_ENABLE_MODULES = YES;
+				CLANG_ENABLE_OBJC_ARC = YES;
+				CLANG_ENABLE_OBJC_WEAK = YES;
+				CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES;
+				CLANG_WARN_BOOL_CONVERSION = YES;
+				CLANG_WARN_COMMA = YES;
+				CLANG_WARN_CONSTANT_CONVERSION = YES;
+				CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES;
+				CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR;
+				CLANG_WARN_DOCUMENTATION_COMMENTS = YES;
+				CLANG_WARN_EMPTY_BODY = YES;
+				CLANG_WARN_ENUM_CONVERSION = YES;
+				CLANG_WARN_INFINITE_RECURSION = YES;
+				CLANG_WARN_INT_CONVERSION = YES;
+				CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES;
+				CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES;
+				CLANG_WARN_OBJC_LITERAL_CONVERSION = YES;
+				CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR;
+				CLANG_WARN_QUOTED_INCLUDE_IN_FRAMEWORK_HEADER = YES;
+				CLANG_WARN_RANGE_LOOP_ANALYSIS = YES;
+				CLANG_WARN_STRICT_PROTOTYPES = YES;
+				CLANG_WARN_SUSPICIOUS_MOVE = YES;
+				CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE;
+				CLANG_WARN_UNREACHABLE_CODE = YES;
+				CLANG_WARN__DUPLICATE_METHOD_MATCH = YES;
+				COPY_PHASE_STRIP = NO;
+				DEBUG_INFORMATION_FORMAT = "dwarf-with-dsym";
+				ENABLE_NS_ASSERTIONS = NO;
+				ENABLE_STRICT_OBJC_MSGSEND = YES;
+				GCC_C_LANGUAGE_STANDARD = gnu11;
+				GCC_NO_COMMON_BLOCKS = YES;
+				GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
+				GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR;
+				GCC_WARN_UNDECLARED_SELECTOR = YES;
+				GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
+				GCC_WARN_UNUSED_FUNCTION = YES;
+				GCC_WARN_UNUSED_VARIABLE = YES;
+				INFOPLIST_KEY_NSMicrophoneUsageDescription = "Needed to transcribe audio";
+				MTL_ENABLE_DEBUG_INFO = NO;
+				MTL_FAST_MATH = YES;
+				SWIFT_COMPILATION_MODE = wholemodule;
+				SWIFT_OPTIMIZATION_LEVEL = "-O";
+			};
+			name = Release;
+		};
+		0AAC5DBD29539CD0003032C3 /* Debug */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
+				ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
+				CLANG_ENABLE_MODULES = YES;
+				CODE_SIGN_ENTITLEMENTS = "whisper.swiftui.demo/Supporting files/WhisperCppDemo.entitlements";
+				CODE_SIGN_STYLE = Automatic;
+				CURRENT_PROJECT_VERSION = 1;
+				DEVELOPMENT_ASSET_PATHS = "\"whisper.swiftui.demo/Supporting files/Preview Content\"";
+				DEVELOPMENT_TEAM = 3TZ9BM962G;
+				ENABLE_HARDENED_RUNTIME = YES;
+				ENABLE_PREVIEWS = YES;
+				GENERATE_INFOPLIST_FILE = YES;
+				"INFOPLIST_KEY_UIApplicationSceneManifest_Generation[sdk=iphoneos*]" = YES;
+				"INFOPLIST_KEY_UIApplicationSceneManifest_Generation[sdk=iphonesimulator*]" = YES;
+				"INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents[sdk=iphoneos*]" = YES;
+				"INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents[sdk=iphonesimulator*]" = YES;
+				"INFOPLIST_KEY_UILaunchScreen_Generation[sdk=iphoneos*]" = YES;
+				"INFOPLIST_KEY_UILaunchScreen_Generation[sdk=iphonesimulator*]" = YES;
+				"INFOPLIST_KEY_UIStatusBarStyle[sdk=iphoneos*]" = UIStatusBarStyleDefault;
+				"INFOPLIST_KEY_UIStatusBarStyle[sdk=iphonesimulator*]" = UIStatusBarStyleDefault;
+				INFOPLIST_KEY_UISupportedInterfaceOrientations_iPad = "UIInterfaceOrientationPortrait UIInterfaceOrientationPortraitUpsideDown UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
+				INFOPLIST_KEY_UISupportedInterfaceOrientations_iPhone = "UIInterfaceOrientationPortrait UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
+				IPHONEOS_DEPLOYMENT_TARGET = 16.1;
+				LD_RUNPATH_SEARCH_PATHS = "@executable_path/Frameworks";
+				"LD_RUNPATH_SEARCH_PATHS[sdk=macosx*]" = "@executable_path/../Frameworks";
+				MACOSX_DEPLOYMENT_TARGET = 13.0;
+				MARKETING_VERSION = 1.0;
+				PRODUCT_BUNDLE_IDENTIFIER = com.whispercppdemo.WhisperCppDemo;
+				PRODUCT_NAME = "$(TARGET_NAME)";
+				SDKROOT = auto;
+				SUPPORTED_PLATFORMS = "iphoneos iphonesimulator macosx";
+				SWIFT_EMIT_LOC_STRINGS = YES;
+				SWIFT_OBJC_BRIDGING_HEADER = "whisper.cpp.swift/WhisperCppDemo-Bridging-Header.h";
+				SWIFT_OPTIMIZATION_LEVEL = "-Onone";
+				SWIFT_VERSION = 5.0;
+				TARGETED_DEVICE_FAMILY = "1,2";
+			};
+			name = Debug;
+		};
+		0AAC5DBE29539CD0003032C3 /* Release */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
+				ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
+				CLANG_ENABLE_MODULES = YES;
+				CODE_SIGN_ENTITLEMENTS = "whisper.swiftui.demo/Supporting files/WhisperCppDemo.entitlements";
+				CODE_SIGN_STYLE = Automatic;
+				CURRENT_PROJECT_VERSION = 1;
+				DEVELOPMENT_ASSET_PATHS = "\"whisper.swiftui.demo/Supporting files/Preview Content\"";
+				DEVELOPMENT_TEAM = 3TZ9BM962G;
+				ENABLE_HARDENED_RUNTIME = YES;
+				ENABLE_PREVIEWS = YES;
+				GENERATE_INFOPLIST_FILE = YES;
+				"INFOPLIST_KEY_UIApplicationSceneManifest_Generation[sdk=iphoneos*]" = YES;
+				"INFOPLIST_KEY_UIApplicationSceneManifest_Generation[sdk=iphonesimulator*]" = YES;
+				"INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents[sdk=iphoneos*]" = YES;
+				"INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents[sdk=iphonesimulator*]" = YES;
+				"INFOPLIST_KEY_UILaunchScreen_Generation[sdk=iphoneos*]" = YES;
+				"INFOPLIST_KEY_UILaunchScreen_Generation[sdk=iphonesimulator*]" = YES;
+				"INFOPLIST_KEY_UIStatusBarStyle[sdk=iphoneos*]" = UIStatusBarStyleDefault;
+				"INFOPLIST_KEY_UIStatusBarStyle[sdk=iphonesimulator*]" = UIStatusBarStyleDefault;
+				INFOPLIST_KEY_UISupportedInterfaceOrientations_iPad = "UIInterfaceOrientationPortrait UIInterfaceOrientationPortraitUpsideDown UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
+				INFOPLIST_KEY_UISupportedInterfaceOrientations_iPhone = "UIInterfaceOrientationPortrait UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
+				IPHONEOS_DEPLOYMENT_TARGET = 16.1;
+				LD_RUNPATH_SEARCH_PATHS = "@executable_path/Frameworks";
+				"LD_RUNPATH_SEARCH_PATHS[sdk=macosx*]" = "@executable_path/../Frameworks";
+				LLVM_LTO = YES;
+				MACOSX_DEPLOYMENT_TARGET = 13.0;
+				MARKETING_VERSION = 1.0;
+				PRODUCT_BUNDLE_IDENTIFIER = com.whispercppdemo.WhisperCppDemo;
+				PRODUCT_NAME = "$(TARGET_NAME)";
+				SDKROOT = auto;
+				SUPPORTED_PLATFORMS = "iphoneos iphonesimulator macosx";
+				SWIFT_EMIT_LOC_STRINGS = YES;
+				SWIFT_OBJC_BRIDGING_HEADER = "whisper.cpp.swift/WhisperCppDemo-Bridging-Header.h";
+				SWIFT_VERSION = 5.0;
+				TARGETED_DEVICE_FAMILY = "1,2";
+			};
+			name = Release;
+		};
+/* End XCBuildConfiguration section */
+
+/* Begin XCConfigurationList section */
+		0AAC5D9229539CCF003032C3 /* Build configuration list for PBXProject "whisper.swiftui" */ = {
+			isa = XCConfigurationList;
+			buildConfigurations = (
+				0AAC5DBA29539CD0003032C3 /* Debug */,
+				0AAC5DBB29539CD0003032C3 /* Release */,
+			);
+			defaultConfigurationIsVisible = 0;
+			defaultConfigurationName = Release;
+		};
+		0AAC5DBC29539CD0003032C3 /* Build configuration list for PBXNativeTarget "whisper.swiftui" */ = {
+			isa = XCConfigurationList;
+			buildConfigurations = (
+				0AAC5DBD29539CD0003032C3 /* Debug */,
+				0AAC5DBE29539CD0003032C3 /* Release */,
+			);
+			defaultConfigurationIsVisible = 0;
+			defaultConfigurationName = Release;
+		};
+/* End XCConfigurationList section */
+	};
+	rootObject = 0AAC5D8F29539CCF003032C3 /* Project object */;
+}

examples/whisper.swiftui/whisper.swiftui.xcodeproj/project.xcworkspace/.gitignore

@@ -0,0 +1 @@
+contents.xcworkspacedata

examples/whisper.swiftui/whisper.swiftui.xcodeproj/project.xcworkspace/xcshareddata/IDEWorkspaceChecks.plist

@@ -0,0 +1,8 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+	<key>IDEDidComputeMac32BitWarning</key>
+	<true/>
+</dict>
+</plist>

examples/whisper.swiftui/whisper.swiftui.xcodeproj/xcshareddata/xcschemes/WhisperCppDemo.xcscheme

@@ -0,0 +1,100 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<Scheme
+   LastUpgradeVersion = "1410"
+   version = "1.3">
+   <BuildAction
+      parallelizeBuildables = "YES"
+      buildImplicitDependencies = "YES">
+      <BuildActionEntries>
+         <BuildActionEntry
+            buildForTesting = "YES"
+            buildForRunning = "YES"
+            buildForProfiling = "YES"
+            buildForArchiving = "YES"
+            buildForAnalyzing = "YES">
+            <BuildableReference
+               BuildableIdentifier = "primary"
+               BlueprintIdentifier = "0AAC5D9629539CCF003032C3"
+               BuildableName = "whisper.swiftui.app"
+               BlueprintName = "whisper.swiftui"
+               ReferencedContainer = "container:whisper.swiftui.xcodeproj">
+            </BuildableReference>
+         </BuildActionEntry>
+      </BuildActionEntries>
+   </BuildAction>
+   <TestAction
+      buildConfiguration = "Debug"
+      selectedDebuggerIdentifier = "Xcode.DebuggerFoundation.Debugger.LLDB"
+      selectedLauncherIdentifier = "Xcode.DebuggerFoundation.Launcher.LLDB"
+      shouldUseLaunchSchemeArgsEnv = "YES">
+      <Testables>
+         <TestableReference
+            skipped = "NO"
+            parallelizable = "YES">
+            <BuildableReference
+               BuildableIdentifier = "primary"
+               BlueprintIdentifier = "0AAC5DA729539CD0003032C3"
+               BuildableName = "whisper.swiftuiTests.xctest"
+               BlueprintName = "whisper.swiftuiTests"
+               ReferencedContainer = "container:whisper.swiftui.xcodeproj">
+            </BuildableReference>
+         </TestableReference>
+         <TestableReference
+            skipped = "NO"
+            parallelizable = "YES">
+            <BuildableReference
+               BuildableIdentifier = "primary"
+               BlueprintIdentifier = "0AAC5DB129539CD0003032C3"
+               BuildableName = "whisper.swiftuiUITests.xctest"
+               BlueprintName = "whisper.swiftuiUITests"
+               ReferencedContainer = "container:whisper.swiftui.xcodeproj">
+            </BuildableReference>
+         </TestableReference>
+      </Testables>
+   </TestAction>
+   <LaunchAction
+      buildConfiguration = "Release"
+      selectedDebuggerIdentifier = "Xcode.DebuggerFoundation.Debugger.LLDB"
+      selectedLauncherIdentifier = "Xcode.DebuggerFoundation.Launcher.LLDB"
+      launchStyle = "0"
+      useCustomWorkingDirectory = "NO"
+      ignoresPersistentStateOnLaunch = "NO"
+      debugDocumentVersioning = "YES"
+      debugServiceExtension = "internal"
+      allowLocationSimulation = "YES">
+      <BuildableProductRunnable
+         runnableDebuggingMode = "0">
+         <BuildableReference
+            BuildableIdentifier = "primary"
+            BlueprintIdentifier = "0AAC5D9629539CCF003032C3"
+            BuildableName = "whisper.swiftui.app"
+            BlueprintName = "whisper.swiftui"
+            ReferencedContainer = "container:whisper.swiftui.xcodeproj">
+         </BuildableReference>
+      </BuildableProductRunnable>
+   </LaunchAction>
+   <ProfileAction
+      buildConfiguration = "Release"
+      shouldUseLaunchSchemeArgsEnv = "YES"
+      savedToolIdentifier = ""
+      useCustomWorkingDirectory = "NO"
+      debugDocumentVersioning = "YES">
+      <BuildableProductRunnable
+         runnableDebuggingMode = "0">
+         <BuildableReference
+            BuildableIdentifier = "primary"
+            BlueprintIdentifier = "0AAC5D9629539CCF003032C3"
+            BuildableName = "whisper.swiftui.app"
+            BlueprintName = "whisper.swiftui"
+            ReferencedContainer = "container:whisper.swiftui.xcodeproj">
+         </BuildableReference>
+      </BuildableProductRunnable>
+   </ProfileAction>
+   <AnalyzeAction
+      buildConfiguration = "Debug">
+   </AnalyzeAction>
+   <ArchiveAction
+      buildConfiguration = "Release"
+      revealArchiveInOrganizer = "YES">
+   </ArchiveAction>
+</Scheme>

ggml.h

@@ -177,11 +177,13 @@ extern "C" {
 #include <stddef.h>
 #include <stdbool.h>
 
-#define GGML_MAX_DIMS     4
-#define GGML_MAX_NODES    4096
-#define GGML_MAX_PARAMS   16
-#define GGML_MAX_CONTEXTS 64
-#define GGML_MAX_OPT      4
+#define GGML_MAX_DIMS         4
+#define GGML_MAX_NODES        4096
+#define GGML_MAX_PARAMS       16
+#define GGML_MAX_CONTEXTS     64
+#define GGML_MAX_OPT          4
+#define GGML_MAX_THREADS      64
+#define GGML_MAX_THREAD_POOLS 16
 
 #ifdef __ARM_NEON
 // we use the built-in 16-bit float type
@@ -724,7 +726,9 @@ enum ggml_opt_result ggml_opt(
 int ggml_cpu_has_avx(void);
 int ggml_cpu_has_avx2(void);
 int ggml_cpu_has_avx512(void);
+int ggml_cpu_has_fma(void);
 int ggml_cpu_has_neon(void);
+int ggml_cpu_has_arm_fma(void);
 int ggml_cpu_has_f16c(void);
 int ggml_cpu_has_fp16_va(void);
 int ggml_cpu_has_wasm_simd(void);

models/download-ggml-model.sh

@@ -19,7 +19,7 @@ function get_script_path() {
     fi
 }
 
-models_path=$(get_script_path)
+models_path="$(get_script_path)"
 
 # Whisper models
 models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large" )

whisper.cpp

@@ -621,7 +621,6 @@ static bool whisper_model_load(const std::string & fname, whisper_context & wctx
     const ggml_type wtype = model.hparams.f16 ? GGML_TYPE_F16 : GGML_TYPE_F32;
 
     size_t ctx_size = 0;
-    size_t ctx_mem_size = 0;
 
     {
         const auto & hparams = model.hparams;
@@ -730,12 +729,6 @@ static bool whisper_model_load(const std::string & fname, whisper_context & wctx
             ctx_size += n_text_layer*(             n_text_state*ggml_type_size(GGML_TYPE_F32)); // cross_attn_ln_1_b
         }
 
-        ctx_mem_size += n_text_layer*n_text_ctx*n_text_state*ggml_type_size(GGML_TYPE_F16); // memory_k
-        ctx_mem_size += n_text_layer*n_text_ctx*n_text_state*ggml_type_size(GGML_TYPE_F16); // memory_v
-
-        ctx_mem_size += n_text_layer*n_audio_ctx*n_text_state*ggml_type_size(GGML_TYPE_F16); // memory_cross_k
-        ctx_mem_size += n_text_layer*n_audio_ctx*n_text_state*ggml_type_size(GGML_TYPE_F16); // memory_cross_v
-
         ctx_size += (15 + 15*n_audio_layer + 24*n_text_layer)*256; // object overhead
 
         fprintf(stderr, "%s: ggml ctx size = %7.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
@@ -1031,7 +1024,7 @@ static bool whisper_model_load(const std::string & fname, whisper_context & wctx
             fin.read( &tmp[0], tmp.size() ); // read to buffer
             name.assign(&tmp[0], tmp.size());
 
-            if (model.tensors.find(name.data()) == model.tensors.end()) {
+            if (model.tensors.find(name) == model.tensors.end()) {
                 fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
                 return false;
             }
@@ -1387,8 +1380,8 @@ static bool whisper_encode(
         // input for next layer (inpO -> inpL)
         memcpy(inpL->data, inpO->data, ggml_nbytes(inpL));
         inpL->op = GGML_OP_NONE;
-        inpL->src0 = NULL;
-        inpL->src1 = NULL;
+        inpL->src0 = nullptr;
+        inpL->src1 = nullptr;
 
         //printf("%s: - used_mem(%d) = %f MB\n", __func__, il, ggml_used_mem(ctxL)/1024.0/1024.0);
 
@@ -1441,8 +1434,8 @@ static bool whisper_encode(
 
         // TODO: hack to disconnect the encoded features from the previous graph
         cur->op = GGML_OP_NONE;
-        cur->src0 = NULL;
-        cur->src1 = NULL;
+        cur->src0 = nullptr;
+        cur->src1 = nullptr;
 
         for (int il = 0; il < model.hparams.n_text_layer; ++il) {
             auto & layer = model.layers_decoder[il];
@@ -1799,8 +1792,8 @@ static bool whisper_decode(
         // input for next layer (inpO -> inpL)
         memcpy(inpL->data, inpO->data, ggml_nbytes(inpL));
         inpL->op = GGML_OP_NONE;
-        inpL->src0 = NULL;
-        inpL->src1 = NULL;
+        inpL->src0 = nullptr;
+        inpL->src1 = nullptr;
 
         if (N > 1) {
             //printf("%s: - used_mem(%d) = %f MB\n", __func__, il, ggml_used_mem(ctxL)/1024.0/1024.0);
@@ -1870,7 +1863,7 @@ static whisper_token_data whisper_sample_best(
     probs_id.reserve(n_logits);
 
     for (int i = 0; i < n_logits; i++) {
-        probs_id.push_back(std::make_pair(probs[i], i));
+        probs_id.emplace_back(probs[i], i);
     }
 
     {
@@ -2043,7 +2036,7 @@ static void fft(const std::vector<float> & in, std::vector<float> & out) {
 static bool log_mel_spectrogram(
     const float * samples,
     const int n_samples,
-    const int sample_rate,
+    const int /*sample_rate*/,
     const int fft_size,
     const int fft_step,
     const int n_mel,
@@ -2194,7 +2187,7 @@ static std::vector<whisper_vocab::id> tokenize(const whisper_vocab & vocab, cons
     // find the longest tokens that form the words:
     std::vector<whisper_vocab::id> tokens;
     for (const auto & word : words) {
-        if (word.size() == 0) continue;
+        if (word.empty()) continue;
 
         int i = 0;
         int n = word.size();
@@ -2242,7 +2235,8 @@ struct whisper_context * whisper_init(const char * path_model) {
 
     if (!whisper_model_load(path_model, *ctx)) {
         fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, path_model);
-        return NULL;
+        delete ctx;
+        return nullptr;
     }
 
     ctx->t_load_us = ggml_time_us() - t_start_us;
@@ -2360,12 +2354,12 @@ struct whisper_token_data whisper_sample_timestamp(struct whisper_context * ctx,
 int whisper_tokenize(struct whisper_context * ctx, const char * text, whisper_token * tokens, int n_max_tokens) {
     const auto res = tokenize(ctx->vocab, text);
 
-    if (res.size() > n_max_tokens) {
+    if (n_max_tokens < (int) res.size()) {
         fprintf(stderr, "%s: too many resulting tokens: %d (max %d)\n", __func__, (int) res.size(), n_max_tokens);
         return -1;
     }
 
-    for (int i = 0; i < res.size(); i++) {
+    for (int i = 0; i < (int) res.size(); i++) {
         tokens[i] = res[i];
     }
 
@@ -2404,7 +2398,7 @@ const char * whisper_lang_str(int id) {
     }
 
     fprintf(stderr, "%s: unknown language id %d\n", __func__, id);
-    return NULL;
+    return nullptr;
 }
 
 int whisper_lang_auto_detect(
@@ -2438,9 +2432,9 @@ int whisper_lang_auto_detect(
     }
 
     std::vector<std::pair<float, int>> probs_id;
-    for (const auto kv : g_lang) {
+    for (const auto & kv : g_lang) {
         const auto token_lang = whisper_token_lang(ctx, kv.second.first);
-        probs_id.push_back({ ctx->probs[token_lang], kv.second.first });
+        probs_id.emplace_back( ctx->probs[token_lang], kv.second.first );
     }
 
     // sort descending
@@ -2464,7 +2458,7 @@ int whisper_lang_auto_detect(
     }
 
     {
-        for (int i = 0; i < probs_id.size(); i++) {
+        for (int i = 0; i < (int) probs_id.size(); i++) {
             if (lang_probs) {
                 lang_probs[probs_id[i].second] = probs_id[i].first;
             }
@@ -2561,7 +2555,9 @@ const char * whisper_print_system_info(void) {
     s += "AVX = "       + std::to_string(ggml_cpu_has_avx())       + " | ";
     s += "AVX2 = "      + std::to_string(ggml_cpu_has_avx2())      + " | ";
     s += "AVX512 = "    + std::to_string(ggml_cpu_has_avx512())    + " | ";
+    s += "FMA = "       + std::to_string(ggml_cpu_has_fma())       + " | ";
     s += "NEON = "      + std::to_string(ggml_cpu_has_neon())      + " | ";
+    s += "ARM_FMA = "   + std::to_string(ggml_cpu_has_arm_fma())   + " | ";
     s += "F16C = "      + std::to_string(ggml_cpu_has_f16c())      + " | ";
     s += "FP16_VA = "   + std::to_string(ggml_cpu_has_fp16_va())   + " | ";
     s += "WASM_SIMD = " + std::to_string(ggml_cpu_has_wasm_simd()) + " | ";
@@ -2874,7 +2870,7 @@ int whisper_full(
         prompt.clear();
 
         // if we have already generated some text, use it as a prompt to condition the next generation
-        if (prompt_past.size() > 0) {
+        if (!prompt_past.empty()) {
             int n_take = std::min(std::min(params.n_max_text_ctx, whisper_n_text_ctx(ctx)/2), int(prompt_past.size()));
 
             prompt = { whisper_token_prev(ctx) };
@@ -2985,7 +2981,7 @@ int whisper_full(
         if (failed) {
             // when we fail to sample timestamp token, retry by clearing the past prompt
             // if it fails again, then we advance the window by 1 second
-            if (prompt_past.size() > 0) {
+            if (!prompt_past.empty()) {
                 prompt_past.clear();
             } else {
                 fprintf(stderr, "\n%s: failed to generate timestamp token - skipping one second\n\n", __func__);
@@ -3002,11 +2998,11 @@ int whisper_full(
         }
 
         // store the text from this iteration
-        if (tokens_cur.size() > 0) {
+        if (!tokens_cur.empty()) {
             int  i0 = 0;
             auto t0 = seek + 2*(tokens_cur.front().tid - whisper_token_beg(ctx));
 
-            std::string text = "";
+            std::string text;
 
             for (int i = 0; i < (int) tokens_cur.size(); i++) {
                 //printf("%s: %18s %6.3f %18s %6.3f\n", __func__,
@@ -3213,7 +3209,7 @@ int whisper_full_parallel(
             results_i[j].t1 += 100*((i + 1)*n_samples_per_processor)/WHISPER_SAMPLE_RATE + offset_t;
 
             // make sure that segments are not overlapping
-            if (ctx->result_all.size() > 0) {
+            if (!ctx->result_all.empty()) {
                 results_i[j].t0 = std::max(results_i[j].t0, ctx->result_all.back().t1);
             }
 

whisper.h

@@ -148,7 +148,7 @@ extern "C" {
             struct whisper_context * ctx,
                         const char * text,
                      whisper_token * tokens,
-                               int   n_max_tokens);
+	                           int   n_max_tokens);
 
     // Largest language id (i.e. number of available languages - 1)
     WHISPER_API int whisper_lang_max_id();