make : fix MUSL Linux build

* Added whisper state + default state on the whisper_context

* Fixed some examples and bindings

* Fixed whisper_n_len (which was used in some binding) and added whisper_n_len_from_state

* Fixed comments

* whisper : reuse kv_cache_free() and fix compiler warnings

* whisper : clean-up the API comments

---------

Co-authored-by: Sandro Hanea <sandrohanea@microsoft.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

.gitignore

@@ -10,6 +10,7 @@ build-em/
 build-debug/
 build-release/
 build-static/
+build-no-accel/
 build-sanitize-addr/
 build-sanitize-thread/
 

CMakeLists.txt

@@ -1,6 +1,6 @@
 cmake_minimum_required (VERSION 3.0)
 
-project(whisper.cpp VERSION 1.2.0)
+project(whisper.cpp VERSION 1.2.1)
 
 # Add path to modules
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")

Makefile

@@ -30,10 +30,16 @@ endif
 # Compile flags
 #
 
-CFLAGS   = -I.              -O3 -std=c11   -fPIC
-CXXFLAGS = -I. -I./examples -O3 -std=c++11 -fPIC
+CFLAGS   = -I.              -O3 -DNDEBUG -std=c11   -fPIC
+CXXFLAGS = -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC
 LDFLAGS  =
 
+# ref: https://github.com/ggerganov/whisper.cpp/issues/37
+ifneq (,$(findstring musl,$(CCV)))
+	CFLAGS   += -D_POSIX_SOURCE -D_GNU_SOURCE
+	CXXFLAGS += -D_POSIX_SOURCE -D_GNU_SOURCE
+endif
+
 # OS specific
 # TODO: support Windows
 ifeq ($(UNAME_S),Linux)
@@ -141,6 +147,8 @@ ifdef WHISPER_GPROF
 	CXXFLAGS += -pg
 endif
 ifneq ($(filter aarch64%,$(UNAME_M)),)
+	CFLAGS += -mcpu=native
+	CXXFLAGS += -mcpu=native
 endif
 ifneq ($(filter armv6%,$(UNAME_M)),)
 	# Raspberry Pi 1, 2, 3

README.md

@@ -4,7 +4,7 @@
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
 [![npm](https://img.shields.io/npm/v/whisper.cpp.svg)](https://www.npmjs.com/package/whisper.cpp/)
 
-Stable: [v1.2.0](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.2.0) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+Stable: [v1.2.1](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.2.1) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
 
 High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisper) automatic speech recognition (ASR) model:
 
@@ -469,7 +469,9 @@ in [models](models).
 - [X] .NET: | [#422](https://github.com/ggerganov/whisper.cpp/discussions/422)
   - [sandrohanea/whisper.net](https://github.com/sandrohanea/whisper.net)
   - [NickDarvey/whisper](https://github.com/NickDarvey/whisper)
-- [ ] Python: soon | [WIP](https://github.com/ggerganov/whisper.cpp/issues/9)
+- [X] Python: | [#9](https://github.com/ggerganov/whisper.cpp/issues/9)
+  - [stlukey/whispercpp.py](https://github.com/stlukey/whispercpp.py) (Cython)
+  - [aarnphm/whispercpp](https://github.com/aarnphm/whispercpp) (Pybind11)
 
 ## Examples
 

bindings/go/pkg/whisper/model.go

@@ -94,6 +94,7 @@ func (model *model) NewContext() (Context, error) {
 	params.SetPrintRealtime(false)
 	params.SetPrintTimestamps(false)
 	params.SetThreads(runtime.NumCPU())
+	params.SetNoContext(true)
 
 	// Return new context
 	return newContext(model, params)

bindings/go/whisper.go

@@ -20,7 +20,7 @@ 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) {
+static void whisper_new_segment_cb(struct whisper_context* ctx, struct whisper_state* state, int n_new, void* user_data) {
     if(user_data != NULL && ctx != NULL) {
         callNewSegment(user_data, n_new);
     }
@@ -29,7 +29,7 @@ static void whisper_new_segment_cb(struct whisper_context* ctx, int n_new, void*
 // 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) {
+static bool whisper_encoder_begin_cb(struct whisper_context* ctx, struct whisper_state* state, void* user_data) {
     if(user_data != NULL && ctx != NULL) {
         return callEncoderBegin(user_data);
     }

bindings/javascript/package.json

@@ -1,6 +1,6 @@
 {
   "name": "whisper.cpp",
-  "version": "1.2.0",
+  "version": "1.2.1",
   "description": "Whisper speech recognition",
   "main": "whisper.js",
   "scripts": {

bindings/ruby/ext/ruby_whisper.cpp

@@ -199,7 +199,7 @@ static VALUE ruby_whisper_transcribe(int argc, VALUE *argv, VALUE self) {
   {
     static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
 
-    rwp->params.encoder_begin_callback = [](struct whisper_context * /*ctx*/, void * user_data) {
+    rwp->params.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
       bool is_aborted = *(bool*)user_data;
       return !is_aborted;
     };

examples/addon.node/addon.cpp

@@ -72,7 +72,7 @@ int timestamp_to_sample(int64_t t, int n_samples) {
     return std::max(0, std::min((int) n_samples - 1, (int) ((t*WHISPER_SAMPLE_RATE)/100)));
 }
 
-void whisper_print_segment(struct whisper_context * ctx, int n_new, void * user_data) {
+void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * state, int n_new, void * user_data) {
     const auto & params  = *((whisper_print_user_data *) user_data)->params;
     const auto & pcmf32s = *((whisper_print_user_data *) user_data)->pcmf32s;
 
@@ -250,7 +250,7 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
 
             // this callback is called on each new segment
             if (!wparams.print_realtime) {
-                wparams.new_segment_callback           = whisper_print_segment;
+                wparams.new_segment_callback           = whisper_print_segment_callback;
                 wparams.new_segment_callback_user_data = &user_data;
             }
 
@@ -260,7 +260,7 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
             {
                 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*/, struct whisper_state * /*state*/, void * user_data) {
                     bool is_aborted = *(bool*)user_data;
                     return !is_aborted;
                 };

examples/command/command.cpp

@@ -109,73 +109,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "\n");
 }
 
-struct whisper_logits_filter_user_data {
-    std::vector<std::string>                * allowed_commands;
-    std::vector<std::vector<whisper_token>> * allowed_tokens;
-};
-
-void whisper_logits_filter(
-            struct whisper_context * ctx,
-          const whisper_token_data * tokens,
-                               int   n_tokens,
-                             float * logits,
-                              void * user_data){
-    const auto & allowed_tokens = *((whisper_logits_filter_user_data *) user_data)->allowed_tokens;
-
-    printf("n_tokens = %d\n", n_tokens);
-    for (int i = 0; i < n_tokens; i++) {
-        printf(" - '%s' (%.2f)\n", whisper_token_to_str(ctx, tokens[i].id), logits[i]);
-    }
-
-    if (n_tokens == 0) {
-        return;
-    }
-
-    std::vector<std::pair<whisper_token, float>> pool;
-    for (int i = 0; i < (int) allowed_tokens.size(); i++) {
-        const int n = (int) allowed_tokens[i].size();
-        if (n_tokens > n) {
-            continue;
-        }
-
-        const whisper_token id = allowed_tokens[i][n_tokens - 1];
-        pool.push_back({ id, logits[id] });
-    }
-
-    if (pool.empty()) {
-        return;
-    }
-
-    printf("applying logits filter, pool size = %d\n", (int) pool.size());
-
-    const int ibeg = whisper_token_beg(ctx);
-
-    double sum_all = 0.0;
-    for (int i = 0; i < ibeg; ++i) {
-        if (logits[i] == -INFINITY) {
-            continue;
-        }
-        sum_all += logits[i];
-    }
-
-    double sum_pool = 0.0;
-    for (int i = 0; i < (int) pool.size(); ++i) {
-        sum_pool += pool[i].second;
-    }
-
-    printf("sum_all = %.2f, sum_pool = %.2f\n", sum_all, sum_pool);
-
-    for (int i = 0; i < ibeg; ++i) {
-        logits[i] = -INFINITY;
-    }
-
-    for (int i = 0; i < (int) pool.size(); ++i) {
-        //logits[pool[i].first] = pool[i].second / sum_pool * sum_all;
-        logits[pool[i].first] = pool[i].second;
-        printf(" - '%s' (%.2f)\n", whisper_token_to_str(ctx, pool[i].first), logits[pool[i].first]);
-    }
-}
-
 std::string transcribe(whisper_context * ctx, const whisper_params & params, const std::vector<float> & pcmf32, float & prob, int64_t & t_ms) {
     const auto t_start = std::chrono::high_resolution_clock::now();
 
@@ -198,8 +131,6 @@ std::string transcribe(whisper_context * ctx, const whisper_params & params, con
     wparams.audio_ctx        = params.audio_ctx;
     wparams.speed_up         = params.speed_up;
 
-    wparams.temperature_inc  = -1.0f;
-
     if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
         return "";
     }
@@ -403,31 +334,22 @@ int process_command_list(struct whisper_context * ctx, audio_async &audio, const
             wparams.translate        = params.translate;
             wparams.no_context       = true;
             wparams.single_segment   = true;
-            //wparams.max_tokens       = 1;
+            wparams.max_tokens       = 1;
             wparams.language         = params.language.c_str();
             wparams.n_threads        = params.n_threads;
 
             wparams.audio_ctx        = params.audio_ctx;
             wparams.speed_up         = params.speed_up;
 
-            wparams.temperature_inc  = -1.0f;
-
             wparams.prompt_tokens    = k_tokens.data();
             wparams.prompt_n_tokens  = k_tokens.size();
 
-            whisper_logits_filter_user_data user_data = { &allowed_commands, &allowed_tokens };
-
-            wparams.logits_filter_callback           = whisper_logits_filter;
-            wparams.logits_filter_callback_user_data = &user_data;
-
             // run the transformer and a single decoding pass
             if (whisper_full(ctx, wparams, pcmf32_cur.data(), pcmf32_cur.size()) != 0) {
                 fprintf(stderr, "%s: ERROR: whisper_full() failed\n", __func__);
                 break;
             }
 
-            fprintf(stdout, "%s: text - '%s'\n", __func__, whisper_full_get_segment_text(ctx, 0));
-
             // estimate command probability
             // NOTE: not optimal
             {
@@ -514,7 +436,7 @@ int process_command_list(struct whisper_context * ctx, audio_async &audio, const
 
 // always-prompt mode
 // transcribe the voice into text after valid prompt
-int process_always_prompt(struct whisper_context * ctx, audio_async & audio, const whisper_params & params) {
+int always_prompt_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params) {
     bool is_running = true;
     bool ask_prompt = true;
 
@@ -574,7 +496,7 @@ int process_always_prompt(struct whisper_context * ctx, audio_async & audio, con
                 const float sim = similarity(prompt, k_prompt);
 
                 //debug
-                //fprintf(stdout, "command size: %d, sim: %f\n", (int) command.size(), sim);
+                //fprintf(stdout, "command size: %i\n", command_length);
 
                 if ((sim > 0.7f) && (command.size() > 0)) {
                     fprintf(stdout, "%s: Command '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", command.c_str(), "\033[0m", (int) t_ms);
@@ -754,7 +676,7 @@ int main(int argc, char ** argv) {
     if (!params.commands.empty()) {
         ret_val = process_command_list(ctx, audio, params);
     } else if (!params.prompt.empty()) {
-        ret_val = process_always_prompt(ctx, audio, params);
+        ret_val = always_prompt_transcription(ctx, audio, params);
     } else {
         ret_val = process_general_transcription(ctx, audio, params);
     }

examples/common-sdl.h

@@ -1,13 +1,13 @@
 #pragma once
 
+#include <SDL.h>
+#include <SDL_audio.h>
+
 #include <atomic>
 #include <cstdint>
 #include <vector>
 #include <mutex>
 
-#include <SDL.h>
-#include <SDL_audio.h>
-
 //
 // SDL Audio capture
 //

examples/main/main.cpp

@@ -80,6 +80,7 @@ struct whisper_params {
 
     std::string language = "en";
     std::string prompt;
+    std::string font_path = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
     std::string model    = "models/ggml-base.en.bin";
 
     std::vector<std::string> fname_inp = {};
@@ -127,6 +128,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-ovtt" || arg == "--output-vtt")     { params.output_vtt     = true; }
         else if (arg == "-osrt" || arg == "--output-srt")     { params.output_srt     = true; }
         else if (arg == "-owts" || arg == "--output-words")   { params.output_wts     = true; }
+        else if (arg == "-fp"   || arg == "--font-path")      { params.font_path      = argv[++i]; }
         else if (arg == "-ocsv" || arg == "--output-csv")     { params.output_csv     = true; }
         else if (arg == "-of"   || arg == "--output-file")    { params.fname_out.emplace_back(argv[++i]); }
         else if (arg == "-ps"   || arg == "--print-special")  { params.print_special  = true; }
@@ -174,6 +176,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -ovtt,     --output-vtt        [%-7s] output result in a vtt file\n",                    params.output_vtt ? "true" : "false");
     fprintf(stderr, "  -osrt,     --output-srt        [%-7s] output result in a srt file\n",                    params.output_srt ? "true" : "false");
     fprintf(stderr, "  -owts,     --output-words      [%-7s] output script for generating karaoke video\n",     params.output_wts ? "true" : "false");
+    fprintf(stderr, "  -fp,       --font-path         [%-7s] path to a monospace font for karaoke video\n",     params.font_path.c_str());
     fprintf(stderr, "  -ocsv,     --output-csv        [%-7s] output result in a CSV file\n",                    params.output_csv ? "true" : "false");
     fprintf(stderr, "  -of FNAME, --output-file FNAME [%-7s] output file path (without file extension)\n",      "");
     fprintf(stderr, "  -ps,       --print-special     [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
@@ -193,7 +196,7 @@ struct whisper_print_user_data {
     const std::vector<std::vector<float>> * pcmf32s;
 };
 
-void whisper_print_segment(struct whisper_context * ctx, int n_new, void * user_data) {
+void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * /*state*/, int n_new, void * user_data) {
     const auto & params  = *((whisper_print_user_data *) user_data)->params;
     const auto & pcmf32s = *((whisper_print_user_data *) user_data)->pcmf32s;
 
@@ -352,13 +355,14 @@ bool output_csv(struct whisper_context * ctx, const char * fname) {
     fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
 
     const int n_segments = whisper_full_n_segments(ctx);
+    fout << "start,end,text\n";
     for (int i = 0; i < n_segments; ++i) {
         const char * text = whisper_full_get_segment_text(ctx, i);
         const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
         const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
 
         //need to multiply times returned from whisper_full_get_segment_t{0,1}() by 10 to get milliseconds.
-        fout << 10 * t0 << ", " << 10 * t1 << ", \"" << text    << "\"\n";
+        fout << 10 * t0 << "," << 10 * t1 << ",\"" << text    << "\"\n";
     }
 
     return true;
@@ -367,13 +371,18 @@ bool output_csv(struct whisper_context * ctx, const char * fname) {
 // 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);
 
-    // TODO: become parameter
-    static const char * font = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
+    static const char * font = params.font_path.c_str();
+
+    std::ifstream fin(font);
+    if (!fin.is_open()) {
+        fprintf(stderr, "%s: font not found at '%s', please specify a monospace font with -fp\n", __func__, font);
+        return false;
+    }
 
     fout << "#!/bin/bash" << "\n";
     fout << "\n";
@@ -597,7 +606,7 @@ int main(int argc, char ** argv) {
 
             // this callback is called on each new segment
             if (!wparams.print_realtime) {
-                wparams.new_segment_callback           = whisper_print_segment;
+                wparams.new_segment_callback           = whisper_print_segment_callback;
                 wparams.new_segment_callback_user_data = &user_data;
             }
 
@@ -607,7 +616,7 @@ 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*/, struct whisper_state * /*state*/, void * user_data) {
                     bool is_aborted = *(bool*)user_data;
                     return !is_aborted;
                 };

examples/stream/stream.cpp

@@ -288,7 +288,6 @@ int main(int argc, char ** argv) {
             wparams.print_realtime   = false;
             wparams.print_timestamps = !params.no_timestamps;
             wparams.translate        = params.translate;
-            wparams.no_context       = true;
             wparams.single_segment   = !use_vad;
             wparams.max_tokens       = params.max_tokens;
             wparams.language         = params.language.c_str();

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

@@ -74,6 +74,7 @@ private class WhisperLib {
         init {
             Log.d(LOG_TAG, "Primary ABI: ${Build.SUPPORTED_ABIS[0]}")
             var loadVfpv4 = false
+            var loadV8fp16 = false
             if (isArmEabiV7a()) {
                 // armeabi-v7a needs runtime detection support
                 val cpuInfo = cpuInfo()
@@ -84,11 +85,24 @@ private class WhisperLib {
                         loadVfpv4 = true
                     }
                 }
+            } else if (isArmEabiV8a()) {
+                // ARMv8.2a needs runtime detection support
+                val cpuInfo = cpuInfo()
+                cpuInfo?.let {
+                    Log.d(LOG_TAG, "CPU info: $cpuInfo")
+                    if (cpuInfo.contains("fphp")) {
+                        Log.d(LOG_TAG, "CPU supports fp16 arithmetic")
+                        loadV8fp16 = true
+                    }
+                }
             }
 
             if (loadVfpv4) {
                 Log.d(LOG_TAG, "Loading libwhisper_vfpv4.so")
                 System.loadLibrary("whisper_vfpv4")
+            } else if (loadV8fp16) {
+                Log.d(LOG_TAG, "Loading libwhisper_v8fp16_va.so")
+                System.loadLibrary("whisper_v8fp16_va")
             } else {
                 Log.d(LOG_TAG, "Loading libwhisper.so")
                 System.loadLibrary("whisper")
@@ -110,6 +124,10 @@ private fun isArmEabiV7a(): Boolean {
     return Build.SUPPORTED_ABIS[0].equals("armeabi-v7a")
 }
 
+private fun isArmEabiV8a(): Boolean {
+    return Build.SUPPORTED_ABIS[0].equals("arm64-v8a")
+}
+
 private fun cpuInfo(): String? {
     return try {
         File("/proc/cpuinfo").inputStream().bufferedReader().use {

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

@@ -12,4 +12,15 @@ ifeq ($(TARGET_ARCH_ABI),armeabi-v7a)
 	# https://android.googlesource.com/platform/ndk/+/master/sources/android/cpufeatures/cpu-features.h
 	LOCAL_CFLAGS += -mfpu=neon-vfpv4
 	include $(BUILD_SHARED_LIBRARY)
-endif
+endif
+
+ifeq ($(TARGET_ARCH_ABI),arm64-v8a)
+	include $(CLEAR_VARS)
+	LOCAL_MODULE    := libwhisper_v8fp16_va
+	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 += -march=armv8.2-a+fp16
+	include $(BUILD_SHARED_LIBRARY)
+endif
+

extra/bench-wts.sh

@@ -0,0 +1,70 @@
+# Benchmark word-level timestamps for different models
+#
+# This script takes two arguments
+# - an audio file
+# - [optional] path to a font file
+
+# I'm using "/usr/share/fonts/truetype/freefont/FreeMono.ttf" on Ubuntu
+
+if [ -z "$1" ]; then
+    echo "Usage: $0 <audio file> [font file]"
+    exit 1
+fi
+
+#TODO: Make this a command line parameter
+#models="base small large"
+#models="tiny.en tiny base.en base small.en small medium.en medium large-v1 large"
+models="tiny.en base.en small.en medium.en large"
+
+DURATION=$(ffprobe -i $1 -show_entries format=duration -v quiet -of csv="p=0")
+DURATION=$(printf "%.2f" $DURATION)
+echo "Input file duration: ${DURATION}s"
+
+for model in $models; do
+    echo "Running $model"
+    COMMAND="./main -m models/ggml-$model.bin -owts -f $1 -of $1.$model"
+
+    if [ ! -z "$2" ]; then
+        COMMAND="$COMMAND -fp $2"
+    fi
+    #TODO: Surface errors better
+    # TIMEFMT is for zsh, TIMEFORMAT is for bash
+    EXECTIME=$({ TIMEFMT="%E";TIMEFORMAT=%E; time $COMMAND >/dev/null 2>&1; } 2>&1)
+
+    # Slightly different formats between zsh and bash
+    if [ "${EXECTIME: -1}" == "s" ]; then
+        EXECTIME=${EXECTIME::-1}
+    fi
+
+    RATIO=$(echo "$DURATION / $EXECTIME" | bc -l)
+    RATIO=$(printf "%.2f" $RATIO)
+
+    echo "Execution time: ${EXECTIME}s (${RATIO}x realtime)"
+
+    # If the file already exists, delete it
+    if [ -f $1.mp4 ]; then
+        rm $1.mp4
+    fi
+
+    bash $1.$model.wts >/dev/null 2>&1
+    mv $1.mp4 $1.$model.mp4
+
+    ffmpeg -y -f lavfi -i color=c=black:s=1200x50:d=$DURATION -vf "drawtext=fontfile=$2:fontsize=36:x=10:y=(h-text_h)/2:text='ggml-$model - ${EXECTIME}s (${RATIO}x realtime)':fontcolor=lightgrey" $1.$model.info.mp4 >/dev/null 2>&1
+done
+
+COMMAND="ffmpeg -y"
+for model in $models; do
+    COMMAND="$COMMAND -i $1.$model.info.mp4 -i $1.$model.mp4"
+done
+COMMAND="$COMMAND -filter_complex \""
+COUNT=0
+for model in $models; do
+    COMMAND="$COMMAND[${COUNT}:v][$(($COUNT+1)):v]"
+    COUNT=$((COUNT+2))
+done
+COMMAND="$COMMAND vstack=inputs=${COUNT}[v]\" -map \"[v]\" -map 1:a $1.all.mp4 >/dev/null 2>&1"
+
+echo $COMMAND
+
+# Run the command
+eval $COMMAND

whisper.h

@@ -66,6 +66,7 @@ extern "C" {
     //
 
     struct whisper_context;
+    struct whisper_state;
 
     typedef int whisper_token;
 
@@ -101,11 +102,20 @@ extern "C" {
     WHISPER_API struct whisper_context * whisper_init_from_buffer(void * buffer, size_t buffer_size);
     WHISPER_API struct whisper_context * whisper_init(struct whisper_model_loader * loader);
 
-    // Frees all memory allocated by the model.
-    WHISPER_API void whisper_free(struct whisper_context * ctx);
+    // These are the same as the above, but the internal state of the context is not allocated automatically
+    // It is the responsibility of the caller to allocate the state using whisper_init_state() (#523)
+    WHISPER_API struct whisper_context * whisper_init_from_file_no_state(const char * path_model);
+    WHISPER_API struct whisper_context * whisper_init_from_buffer_no_state(void * buffer, size_t buffer_size);
+    WHISPER_API struct whisper_context * whisper_init_no_state(struct whisper_model_loader * loader);
+
+    WHISPER_API struct whisper_state * whisper_init_state(struct whisper_context * ctx);
+
+    // Frees all allocated memory
+    WHISPER_API void whisper_free      (struct whisper_context * ctx);
+    WHISPER_API void whisper_free_state(struct whisper_state * state);
 
     // Convert RAW PCM audio to log mel spectrogram.
-    // The resulting spectrogram is stored inside the provided whisper context.
+    // The resulting spectrogram is stored inside the default state of the provided whisper context.
     // Returns 0 on success
     WHISPER_API int whisper_pcm_to_mel(
             struct whisper_context * ctx,
@@ -113,17 +123,30 @@ extern "C" {
                                int   n_samples,
                                int   n_threads);
 
-    // Convert RAW PCM audio to log mel spectrogram but applies a Phase Vocoder to speed up the audio x2. 
-    // The resulting spectrogram is stored inside the provided whisper context.
+    WHISPER_API int whisper_pcm_to_mel_with_state(
+            struct whisper_context * ctx,
+              struct whisper_state * state,
+                       const float * samples,
+                               int   n_samples,
+                               int   n_threads);
+
+    // Convert RAW PCM audio to log mel spectrogram but applies a Phase Vocoder to speed up the audio x2.
+    // The resulting spectrogram is stored inside the default state of the provided whisper context.
     // Returns 0 on success
     WHISPER_API int whisper_pcm_to_mel_phase_vocoder(
-        struct whisper_context* ctx,
-        const float* samples,
-        int   n_samples,
-        int   n_threads);
+        struct whisper_context * ctx,
+                   const float * samples,
+                           int   n_samples,
+                           int   n_threads);
 
+    WHISPER_API int whisper_pcm_to_mel_phase_vocoder_with_state(
+        struct whisper_context * ctx,
+          struct whisper_state * state,
+                   const float * samples,
+                           int   n_samples,
+                           int   n_threads);
 
-    // This can be used to set a custom log mel spectrogram inside the provided whisper context.
+    // This can be used to set a custom log mel spectrogram inside the default state of 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
     // Returns 0 on success
@@ -133,7 +156,14 @@ extern "C" {
                                int   n_len,
                                int   n_mel);
 
-    // Run the Whisper encoder on the log mel spectrogram stored inside the provided whisper context.
+    WHISPER_API int whisper_set_mel_with_state(
+            struct whisper_context * ctx,
+              struct whisper_state * state,
+                       const float * data,
+                               int   n_len,
+                               int   n_mel);
+
+    // Run the Whisper encoder on the log mel spectrogram stored inside the default state in 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.
     // Returns 0 on success
@@ -142,6 +172,12 @@ extern "C" {
                                int   offset,
                                int   n_threads);
 
+    WHISPER_API int whisper_encode_with_state(
+            struct whisper_context * ctx,
+              struct whisper_state * state,
+                               int   offset,
+                               int   n_threads);
+
     // 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.
@@ -155,6 +191,14 @@ extern "C" {
                                int   n_past,
                                int   n_threads);
 
+    WHISPER_API int whisper_decode_with_state(
+            struct whisper_context * ctx,
+              struct whisper_state * state,
+               const whisper_token * tokens,
+                               int   n_tokens,
+                               int   n_past,
+                               int   n_threads);
+
     // 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, no more than n_max_tokens
@@ -190,17 +234,26 @@ extern "C" {
                                int   n_threads,
                              float * lang_probs);
 
-    WHISPER_API int whisper_n_len          (struct whisper_context * ctx); // mel length
-    WHISPER_API int whisper_n_vocab        (struct whisper_context * ctx);
-    WHISPER_API int whisper_n_text_ctx     (struct whisper_context * ctx);
-    WHISPER_API int whisper_n_audio_ctx    (struct whisper_context * ctx);
-    WHISPER_API int whisper_is_multilingual(struct whisper_context * ctx);
+    WHISPER_API int whisper_lang_auto_detect_with_state(
+            struct whisper_context * ctx,
+              struct whisper_state * state,
+                               int   offset_ms,
+                               int   n_threads,
+                             float * lang_probs);
+
+    WHISPER_API int whisper_n_len           (struct whisper_context * ctx); // mel length
+    WHISPER_API int whisper_n_len_from_state(struct whisper_state * state); // mel length
+    WHISPER_API int whisper_n_vocab         (struct whisper_context * ctx);
+    WHISPER_API int whisper_n_text_ctx      (struct whisper_context * ctx);
+    WHISPER_API int whisper_n_audio_ctx     (struct whisper_context * ctx);
+    WHISPER_API int whisper_is_multilingual (struct whisper_context * ctx);
 
     // Token logits obtained from the last call to whisper_decode()
     // The logits for the last token are stored in the last row
     // Rows: n_tokens
     // Cols: n_vocab
-    WHISPER_API float * whisper_get_logits(struct whisper_context * ctx);
+    WHISPER_API float * whisper_get_logits           (struct whisper_context * ctx);
+    WHISPER_API float * whisper_get_logits_from_state(struct whisper_state * state);
 
     // Token Id -> String. Uses the vocabulary in the provided context
     WHISPER_API const char * whisper_token_to_str(struct whisper_context * ctx, whisper_token token);
@@ -218,7 +271,7 @@ extern "C" {
     WHISPER_API whisper_token whisper_token_translate (void);
     WHISPER_API whisper_token whisper_token_transcribe(void);
 
-    // Performance information
+    // Performance information from the default state.
     WHISPER_API void whisper_print_timings(struct whisper_context * ctx);
     WHISPER_API void whisper_reset_timings(struct whisper_context * ctx);
 
@@ -236,18 +289,19 @@ extern "C" {
     // Text segment callback
     // Called on every newly generated text segment
     // Use the whisper_full_...() functions to obtain the text segments
-    typedef void (*whisper_new_segment_callback)(struct whisper_context * ctx, int n_new, void * user_data);
+    typedef void (*whisper_new_segment_callback)(struct whisper_context * ctx, struct whisper_state * state, int n_new, void * user_data);
 
     // Encoder begin callback
     // If not NULL, called before the encoder starts
     // If it returns false, the computation is aborted
-    typedef bool (*whisper_encoder_begin_callback)(struct whisper_context * ctx, void * user_data);
+    typedef bool (*whisper_encoder_begin_callback)(struct whisper_context * ctx, struct whisper_state * state, void * user_data);
 
     // Logits filter callback
     // Can be used to modify the logits before sampling
     // If not NULL, called after applying temperature to logits
     typedef void (*whisper_logits_filter_callback)(
             struct whisper_context * ctx,
+              struct whisper_state * state,
           const whisper_token_data * tokens,
                                int   n_tokens,
                              float * logits,
@@ -334,6 +388,7 @@ extern "C" {
     WHISPER_API struct whisper_full_params whisper_full_default_params(enum whisper_sampling_strategy strategy);
 
     // Run the entire model: PCM -> log mel spectrogram -> encoder -> decoder -> text
+    // Not thread safe for same context
     // Uses the specified decoding strategy to obtain the text.
     WHISPER_API int whisper_full(
                 struct whisper_context * ctx,
@@ -341,7 +396,16 @@ extern "C" {
                            const float * samples,
                                    int   n_samples);
 
-    // Split the input audio in chunks and process each chunk separately using whisper_full()
+    WHISPER_API int whisper_full_with_state(
+                struct whisper_context * ctx,
+                  struct whisper_state * state,
+            struct whisper_full_params   params,
+                           const float * samples,
+                                   int   n_samples);
+
+    // Split the input audio in chunks and process each chunk separately using whisper_full_with_state()
+    // Result is stored in the default state of the context
+    // Not thread safe if executed in parallel on the same context.
     // 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.
     WHISPER_API int whisper_full_parallel(
@@ -351,33 +415,47 @@ extern "C" {
                                    int   n_samples,
                                    int   n_processors);
 
-    // Number of generated text segments.
+    // Number of generated text segments
     // A segment can be a few words, a sentence, or even a paragraph.
-    WHISPER_API int whisper_full_n_segments(struct whisper_context * ctx);
+    WHISPER_API int whisper_full_n_segments           (struct whisper_context * ctx);
+    WHISPER_API int whisper_full_n_segments_from_state(struct whisper_state * state);
 
-    // Language id associated with the current context
+    // Language id associated with the context's default state
     WHISPER_API int whisper_full_lang_id(struct whisper_context * ctx);
 
-    // Get the start and end time of the specified segment.
-    WHISPER_API int64_t whisper_full_get_segment_t0(struct whisper_context * ctx, int i_segment);
-    WHISPER_API int64_t whisper_full_get_segment_t1(struct whisper_context * ctx, int i_segment);
+    // Language id associated with the provided state
+    WHISPER_API int whisper_full_lang_id_from_state(struct whisper_state * state);
 
-    // Get the text of the specified segment.
-    WHISPER_API const char * whisper_full_get_segment_text(struct whisper_context * ctx, int i_segment);
+    // Get the start and end time of the specified segment
+    WHISPER_API int64_t whisper_full_get_segment_t0           (struct whisper_context * ctx, int i_segment);
+    WHISPER_API int64_t whisper_full_get_segment_t0_from_state(struct whisper_state * state, int i_segment);
 
-    // Get number of tokens in the specified segment.
-    WHISPER_API int whisper_full_n_tokens(struct whisper_context * ctx, int i_segment);
+    WHISPER_API int64_t whisper_full_get_segment_t1           (struct whisper_context * ctx, int i_segment);
+    WHISPER_API int64_t whisper_full_get_segment_t1_from_state(struct whisper_state * state, int i_segment);
 
-    // Get the token text of the specified token in the specified segment.
-    WHISPER_API const char * whisper_full_get_token_text(struct whisper_context * ctx, int i_segment, int i_token);
-    WHISPER_API whisper_token whisper_full_get_token_id (struct whisper_context * ctx, int i_segment, int i_token);
+    // Get the text of the specified segment
+    WHISPER_API const char * whisper_full_get_segment_text           (struct whisper_context * ctx, int i_segment);
+    WHISPER_API const char * whisper_full_get_segment_text_from_state(struct whisper_state * state, int i_segment);
 
-    // Get token data for the specified token in the specified segment.
+    // Get number of tokens in the specified segment
+    WHISPER_API int whisper_full_n_tokens           (struct whisper_context * ctx, int i_segment);
+    WHISPER_API int whisper_full_n_tokens_from_state(struct whisper_state * state, int i_segment);
+
+    // Get the token text of the specified token in the specified segment
+    WHISPER_API const char * whisper_full_get_token_text           (struct whisper_context * ctx, int i_segment, int i_token);
+    WHISPER_API const char * whisper_full_get_token_text_from_state(struct whisper_context * ctx, struct whisper_state * state, int i_segment, int i_token);
+
+    WHISPER_API whisper_token whisper_full_get_token_id           (struct whisper_context * ctx, int i_segment, int i_token);
+    WHISPER_API whisper_token whisper_full_get_token_id_from_state(struct whisper_state * state, int i_segment, int i_token);
+
+    // Get token data for the specified token in the specified segment
     // This contains probabilities, timestamps, etc.
-    WHISPER_API whisper_token_data whisper_full_get_token_data(struct whisper_context * ctx, int i_segment, int i_token);
+    WHISPER_API whisper_token_data whisper_full_get_token_data           (struct whisper_context * ctx, int i_segment, int i_token);
+    WHISPER_API whisper_token_data whisper_full_get_token_data_from_state(struct whisper_state * state, int i_segment, int i_token);
 
-    // Get the probability of the specified token in the specified segment.
-    WHISPER_API float whisper_full_get_token_p(struct whisper_context * ctx, int i_segment, int i_token);
+    // Get the probability of the specified token in the specified segment
+    WHISPER_API float whisper_full_get_token_p           (struct whisper_context * ctx, int i_segment, int i_token);
+    WHISPER_API float whisper_full_get_token_p_from_state(struct whisper_state * state, int i_segment, int i_token);
 
     ////////////////////////////////////////////////////////////////////////////