mirror of
https://github.com/ggerganov/whisper.cpp.git
synced 2025-07-04 16:30:58 +02:00
Compare commits
2 Commits
coreml-wit
...
coreml
| Author | SHA1 | Date | |
|---|---|---|---|
| 17a14593de | |||
| b0ac915265 |
@ -191,11 +191,9 @@ else()
|
||||
if(NOT WHISPER_NO_FMA)
|
||||
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mfma")
|
||||
endif()
|
||||
if(NOT WHISPER_NO_F16C)
|
||||
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mf16c")
|
||||
endif()
|
||||
endif()
|
||||
endif()
|
||||
endif()
|
||||
|
||||
if (WHISPER_PERF)
|
||||
|
||||
6
Makefile
6
Makefile
@ -34,12 +34,6 @@ 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 ($(wildcard /usr/include/musl/*),)
|
||||
CFLAGS += -D_POSIX_SOURCE -D_GNU_SOURCE
|
||||
CXXFLAGS += -D_POSIX_SOURCE -D_GNU_SOURCE
|
||||
endif
|
||||
|
||||
# OS specific
|
||||
# TODO: support Windows
|
||||
ifeq ($(UNAME_S),Linux)
|
||||
|
||||
17
README.md
17
README.md
@ -433,19 +433,6 @@ https://user-images.githubusercontent.com/1991296/199337538-b7b0c7a3-2753-4a88-a
|
||||
|
||||
---
|
||||
|
||||
## Video comparison of different models
|
||||
|
||||
Use the [extra/bench-wts.sh](https://github.com/ggerganov/whisper.cpp/blob/master/extra/bench-wts.sh) script to generate a video in the following format:
|
||||
|
||||
```java
|
||||
./extra/bench-wts.sh samples/jfk.wav
|
||||
ffplay ./samples/jfk.wav.all.mp4
|
||||
```
|
||||
|
||||
https://user-images.githubusercontent.com/1991296/223206245-2d36d903-cf8e-4f09-8c3b-eb9f9c39d6fc.mp4
|
||||
|
||||
---
|
||||
|
||||
## Benchmarks
|
||||
|
||||
In order to have an objective comparison of the performance of the inference across different system configurations,
|
||||
@ -466,7 +453,7 @@ The original models are converted to a custom binary format. This allows to pack
|
||||
You can download the converted models using the [models/download-ggml-model.sh](models/download-ggml-model.sh) script
|
||||
or manually from here:
|
||||
|
||||
- https://huggingface.co/ggerganov/whisper.cpp
|
||||
- https://huggingface.co/datasets/ggerganov/whisper.cpp
|
||||
- https://ggml.ggerganov.com
|
||||
|
||||
For more details, see the conversion script [models/convert-pt-to-ggml.py](models/convert-pt-to-ggml.py) or the README
|
||||
@ -476,7 +463,6 @@ in [models](models).
|
||||
|
||||
- [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)
|
||||
- React Native (iOS / Android): [whisper.rn](https://github.com/mybigday/whisper.rn)
|
||||
- [X] Go: [bindings/go](bindings/go) | [#312](https://github.com/ggerganov/whisper.cpp/discussions/312)
|
||||
- [X] Ruby: [bindings/ruby](bindings/ruby) | [#507](https://github.com/ggerganov/whisper.cpp/discussions/507)
|
||||
- [X] Objective-C / Swift: [ggerganov/whisper.spm](https://github.com/ggerganov/whisper.spm) | [#313](https://github.com/ggerganov/whisper.cpp/discussions/313)
|
||||
@ -486,7 +472,6 @@ in [models](models).
|
||||
- [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)
|
||||
- [X] R: [bnosac/audio.whisper](https://github.com/bnosac/audio.whisper)
|
||||
|
||||
## Examples
|
||||
|
||||
|
||||
@ -17,7 +17,7 @@ import (
|
||||
// CONSTANTS
|
||||
|
||||
const (
|
||||
srcUrl = "https://huggingface.co/ggerganov/whisper.cpp/resolve/main" // The location of the models
|
||||
srcUrl = "https://huggingface.co/datasets/ggerganov/whisper.cpp/resolve/main" // The location of the models
|
||||
srcExt = ".bin" // Filename extension
|
||||
bufSize = 1024 * 64 // Size of the buffer used for downloading the model
|
||||
)
|
||||
|
||||
@ -94,7 +94,6 @@ 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)
|
||||
|
||||
@ -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, struct whisper_state* state, int n_new, void* user_data) {
|
||||
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);
|
||||
}
|
||||
@ -29,7 +29,7 @@ static void whisper_new_segment_cb(struct whisper_context* ctx, struct whisper_s
|
||||
// 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, struct whisper_state* state, void* user_data) {
|
||||
static bool whisper_encoder_begin_cb(struct whisper_context* ctx, void* user_data) {
|
||||
if(user_data != NULL && ctx != NULL) {
|
||||
return callEncoderBegin(user_data);
|
||||
}
|
||||
|
||||
@ -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*/, struct whisper_state * /*state*/, void * user_data) {
|
||||
rwp->params.encoder_begin_callback = [](struct whisper_context * /*ctx*/, void * user_data) {
|
||||
bool is_aborted = *(bool*)user_data;
|
||||
return !is_aborted;
|
||||
};
|
||||
|
||||
@ -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_callback(struct whisper_context * ctx, struct whisper_state * state, int n_new, void * user_data) {
|
||||
void whisper_print_segment_callback(struct whisper_context * ctx, 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;
|
||||
|
||||
@ -260,7 +260,7 @@ int run(whisper_params ¶ms, 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*/, struct whisper_state * /*state*/, void * user_data) {
|
||||
wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, void * user_data) {
|
||||
bool is_aborted = *(bool*)user_data;
|
||||
return !is_aborted;
|
||||
};
|
||||
@ -292,41 +292,13 @@ int run(whisper_params ¶ms, std::vector<std::vector<std::string>> &result) {
|
||||
return 0;
|
||||
}
|
||||
|
||||
class Worker : public Napi::AsyncWorker {
|
||||
public:
|
||||
Worker(Napi::Function& callback, whisper_params params)
|
||||
: Napi::AsyncWorker(callback), params(params) {}
|
||||
|
||||
void Execute() override {
|
||||
run(params, result);
|
||||
}
|
||||
|
||||
void OnOK() override {
|
||||
Napi::HandleScope scope(Env());
|
||||
Napi::Object res = Napi::Array::New(Env(), result.size());
|
||||
for (uint64_t i = 0; i < result.size(); ++i) {
|
||||
Napi::Object tmp = Napi::Array::New(Env(), 3);
|
||||
for (uint64_t j = 0; j < 3; ++j) {
|
||||
tmp[j] = Napi::String::New(Env(), result[i][j]);
|
||||
}
|
||||
res[i] = tmp;
|
||||
}
|
||||
Callback().Call({Env().Null(), res});
|
||||
}
|
||||
|
||||
private:
|
||||
whisper_params params;
|
||||
std::vector<std::vector<std::string>> result;
|
||||
};
|
||||
|
||||
|
||||
|
||||
Napi::Value whisper(const Napi::CallbackInfo& info) {
|
||||
Napi::Object whisper(const Napi::CallbackInfo& info) {
|
||||
Napi::Env env = info.Env();
|
||||
if (info.Length() <= 0 || !info[0].IsObject()) {
|
||||
Napi::TypeError::New(env, "object expected").ThrowAsJavaScriptException();
|
||||
}
|
||||
whisper_params params;
|
||||
std::vector<std::vector<std::string>> result;
|
||||
|
||||
Napi::Object whisper_params = info[0].As<Napi::Object>();
|
||||
std::string language = whisper_params.Get("language").As<Napi::String>();
|
||||
@ -337,10 +309,25 @@ Napi::Value whisper(const Napi::CallbackInfo& info) {
|
||||
params.model = model;
|
||||
params.fname_inp.emplace_back(input);
|
||||
|
||||
Napi::Function callback = info[1].As<Napi::Function>();
|
||||
Worker* worker = new Worker(callback, params);
|
||||
worker->Queue();
|
||||
return env.Undefined();
|
||||
// run model
|
||||
run(params, result);
|
||||
|
||||
fprintf(stderr, "RESULT:\n");
|
||||
for (auto sentence:result) {
|
||||
fprintf(stderr, "t0: %s, t1: %s, content: %s \n",
|
||||
sentence[0].c_str(), sentence[1].c_str(), sentence[2].c_str());
|
||||
}
|
||||
|
||||
Napi::Object res = Napi::Array::New(env, result.size());
|
||||
for (uint64_t i = 0; i < result.size(); ++i) {
|
||||
Napi::Object tmp = Napi::Array::New(env, 3);
|
||||
for (uint64_t j = 0; j < 3; ++j) {
|
||||
tmp[j] = Napi::String::New(env, result[i][j]);
|
||||
}
|
||||
res[i] = tmp;
|
||||
}
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
|
||||
|
||||
@ -1,16 +1,10 @@
|
||||
const path = require("path");
|
||||
const { whisper } = require(path.join(
|
||||
__dirname,
|
||||
"../../build/Release/whisper-addon"
|
||||
));
|
||||
const { promisify } = require("util");
|
||||
|
||||
const whisperAsync = promisify(whisper);
|
||||
const path = require('path');
|
||||
const { whisper } = require(path.join(__dirname, '../../build/Release/whisper-addon'));
|
||||
|
||||
const whisperParams = {
|
||||
language: "en",
|
||||
model: path.join(__dirname, "../../models/ggml-base.en.bin"),
|
||||
fname_inp: "../../samples/jfk.wav",
|
||||
language: 'en',
|
||||
model: path.join(__dirname, '../../models/ggml-base.en.bin'),
|
||||
fname_inp: '',
|
||||
};
|
||||
|
||||
const arguments = process.argv.slice(2);
|
||||
@ -20,7 +14,7 @@ const params = Object.fromEntries(
|
||||
return [...pre, item.slice(2).split("=")];
|
||||
}
|
||||
return pre;
|
||||
}, [])
|
||||
}, []),
|
||||
);
|
||||
|
||||
for (const key in params) {
|
||||
@ -29,8 +23,5 @@ for (const key in params) {
|
||||
}
|
||||
}
|
||||
|
||||
console.log("whisperParams =", whisperParams);
|
||||
|
||||
whisperAsync(whisperParams).then((result) => {
|
||||
console.log(`Result from whisper: ${result}`);
|
||||
});
|
||||
console.log('whisperParams =', whisperParams);
|
||||
console.log(whisper(whisperParams));
|
||||
|
||||
@ -31,7 +31,6 @@ options:
|
||||
-osrt, --output-srt [false ] output result in a srt file
|
||||
-owts, --output-words [false ] output script for generating karaoke video
|
||||
-ocsv, --output-csv [false ] output result in a CSV file
|
||||
-oj, --output-json [false ] output result in a JSON file
|
||||
-of FNAME, --output-file FNAME [ ] output file path (without file extension)
|
||||
-ps, --print-special [false ] print special tokens
|
||||
-pc, --print-colors [false ] print colors
|
||||
|
||||
@ -73,7 +73,6 @@ struct whisper_params {
|
||||
bool output_srt = false;
|
||||
bool output_wts = false;
|
||||
bool output_csv = false;
|
||||
bool output_jsn = false;
|
||||
bool print_special = false;
|
||||
bool print_colors = false;
|
||||
bool print_progress = false;
|
||||
@ -81,7 +80,6 @@ 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 = {};
|
||||
@ -129,9 +127,7 @@ 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 == "-oj" || arg == "--output-json") { params.output_jsn = 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; }
|
||||
else if (arg == "-pc" || arg == "--print-colors") { params.print_colors = true; }
|
||||
@ -178,9 +174,7 @@ 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, " -oj, --output-json [%-7s] output result in a JSON file\n", params.output_jsn ? "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");
|
||||
fprintf(stderr, " -pc, --print-colors [%-7s] print colors\n", params.print_colors ? "true" : "false");
|
||||
@ -199,7 +193,7 @@ struct whisper_print_user_data {
|
||||
const std::vector<std::vector<float>> * pcmf32s;
|
||||
};
|
||||
|
||||
void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * /*state*/, int n_new, void * user_data) {
|
||||
void whisper_print_segment_callback(struct whisper_context * ctx, 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;
|
||||
|
||||
@ -371,144 +365,16 @@ bool output_csv(struct whisper_context * ctx, const char * fname) {
|
||||
return true;
|
||||
}
|
||||
|
||||
bool output_json(struct whisper_context * ctx, const char * fname, const whisper_params & params) {
|
||||
std::ofstream fout(fname);
|
||||
int indent = 0;
|
||||
|
||||
auto doindent = [&]() {
|
||||
for (int i = 0; i < indent; i++) fout << "\t";
|
||||
};
|
||||
|
||||
auto start_arr = [&](const char *name) {
|
||||
doindent();
|
||||
fout << "\"" << name << "\": [\n";
|
||||
indent++;
|
||||
};
|
||||
|
||||
auto end_arr = [&](bool end = false) {
|
||||
indent--;
|
||||
doindent();
|
||||
fout << (end ? "]\n" : "},\n");
|
||||
};
|
||||
|
||||
auto start_obj = [&](const char *name = nullptr) {
|
||||
doindent();
|
||||
if (name) {
|
||||
fout << "\"" << name << "\": {\n";
|
||||
} else {
|
||||
fout << "{\n";
|
||||
}
|
||||
indent++;
|
||||
};
|
||||
|
||||
auto end_obj = [&](bool end = false) {
|
||||
indent--;
|
||||
doindent();
|
||||
fout << (end ? "}\n" : "},\n");
|
||||
};
|
||||
|
||||
auto start_value = [&](const char *name) {
|
||||
doindent();
|
||||
fout << "\"" << name << "\": ";
|
||||
};
|
||||
|
||||
auto value_s = [&](const char *name, const char *val, bool end = false) {
|
||||
start_value(name);
|
||||
fout << "\"" << val << (end ? "\"\n" : "\",\n");
|
||||
};
|
||||
|
||||
auto end_value = [&](bool end = false) {
|
||||
fout << (end ? "\n" : ",\n");
|
||||
};
|
||||
|
||||
auto value_i = [&](const char *name, const int64_t val, bool end = false) {
|
||||
start_value(name);
|
||||
fout << val;
|
||||
end_value(end);
|
||||
};
|
||||
|
||||
auto value_b = [&](const char *name, const bool val, bool end = false) {
|
||||
start_value(name);
|
||||
fout << (val ? "true" : "false");
|
||||
end_value(end);
|
||||
};
|
||||
|
||||
if (!fout.is_open()) {
|
||||
fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
|
||||
return false;
|
||||
}
|
||||
|
||||
fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
|
||||
start_obj();
|
||||
value_s("systeminfo", whisper_print_system_info());
|
||||
start_obj("model");
|
||||
value_s("type", whisper_model_type_readable(ctx));
|
||||
value_b("multilingual", whisper_is_multilingual(ctx));
|
||||
value_i("vocab", whisper_model_n_vocab(ctx));
|
||||
start_obj("audio");
|
||||
value_i("ctx", whisper_model_n_audio_ctx(ctx));
|
||||
value_i("state", whisper_model_n_audio_state(ctx));
|
||||
value_i("head", whisper_model_n_audio_head(ctx));
|
||||
value_i("layer", whisper_model_n_audio_layer(ctx), true);
|
||||
end_obj();
|
||||
start_obj("text");
|
||||
value_i("ctx", whisper_model_n_text_ctx(ctx));
|
||||
value_i("state", whisper_model_n_text_state(ctx));
|
||||
value_i("head", whisper_model_n_text_head(ctx));
|
||||
value_i("leyer", whisper_model_n_text_layer(ctx), true);
|
||||
end_obj();
|
||||
value_i("mels", whisper_model_n_mels(ctx));
|
||||
value_i("f16", whisper_model_f16(ctx), true);
|
||||
end_obj();
|
||||
start_obj("params");
|
||||
value_s("model", params.model.c_str());
|
||||
value_s("language", params.language.c_str());
|
||||
value_b("translate", params.translate, true);
|
||||
end_obj();
|
||||
start_obj("result");
|
||||
value_s("language", whisper_lang_str(whisper_full_lang_id(ctx)), true);
|
||||
end_obj();
|
||||
start_arr("transcription");
|
||||
|
||||
const int n_segments = whisper_full_n_segments(ctx);
|
||||
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);
|
||||
|
||||
start_obj();
|
||||
start_obj("timestanps");
|
||||
value_s("from", to_timestamp(t0, true).c_str());
|
||||
value_s("to", to_timestamp(t1, true).c_str(), true);
|
||||
end_obj();
|
||||
start_obj("offsets");
|
||||
value_i("from", t0 * 10);
|
||||
value_i("to", t1 * 10, true);
|
||||
end_obj();
|
||||
value_s("text", text, true);
|
||||
end_obj(i == (n_segments - 1));
|
||||
}
|
||||
|
||||
end_arr(true);
|
||||
end_obj(true);
|
||||
return true;
|
||||
}
|
||||
|
||||
// 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);
|
||||
|
||||
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;
|
||||
}
|
||||
// TODO: become parameter
|
||||
static const char * font = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
|
||||
|
||||
fout << "#!/bin/bash" << "\n";
|
||||
fout << "\n";
|
||||
@ -742,7 +608,7 @@ 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*/, struct whisper_state * /*state*/, void * user_data) {
|
||||
wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, void * user_data) {
|
||||
bool is_aborted = *(bool*)user_data;
|
||||
return !is_aborted;
|
||||
};
|
||||
@ -788,12 +654,6 @@ int main(int argc, char ** argv) {
|
||||
const auto fname_csv = fname_out + ".csv";
|
||||
output_csv(ctx, fname_csv.c_str());
|
||||
}
|
||||
|
||||
// output to JSON file
|
||||
if (params.output_jsn) {
|
||||
const auto fname_jsn = fname_out + ".json";
|
||||
output_json(ctx, fname_jsn.c_str(), params);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
@ -288,6 +288,7 @@ 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();
|
||||
|
||||
@ -31,7 +31,7 @@ To run this, you will need a ggml GPT-2 model: [instructions](https://github.com
|
||||
Alternatively, you can simply download the smallest ggml GPT-2 117M model (240 MB) like this:
|
||||
|
||||
```
|
||||
wget --quiet --show-progress -O models/ggml-gpt-2-117M.bin https://huggingface.co/ggerganov/ggml/raw/main/ggml-model-gpt-2-117M.bin
|
||||
wget --quiet --show-progress -O models/ggml-gpt-2-117M.bin https://huggingface.co/datasets/ggerganov/ggml/raw/main/ggml-model-gpt-2-117M.bin
|
||||
```
|
||||
|
||||
## TTS
|
||||
|
||||
@ -9,4 +9,4 @@ To use:
|
||||
5. Select the "release" active build variant, and use Android Studio to run and deploy to your device.
|
||||
[^1]: I recommend the tiny or base models for running on an Android device.
|
||||
|
||||
<img width="300" alt="image" src="https://user-images.githubusercontent.com/1670775/221613663-a17bf770-27ef-45ab-9a46-a5f99ba65d2a.jpg">
|
||||
<img width="300" alt="image" src="https://user-images.githubusercontent.com/1991296/208154256-82d972dc-221b-48c4-bfcb-36ce68602f93.png">
|
||||
|
||||
@ -2,7 +2,6 @@ package com.whispercppdemo.ui.main
|
||||
|
||||
import androidx.compose.foundation.layout.*
|
||||
import androidx.compose.foundation.rememberScrollState
|
||||
import androidx.compose.foundation.text.selection.SelectionContainer
|
||||
import androidx.compose.foundation.verticalScroll
|
||||
import androidx.compose.material3.*
|
||||
import androidx.compose.runtime.Composable
|
||||
@ -20,7 +19,6 @@ fun MainScreen(viewModel: MainScreenViewModel) {
|
||||
canTranscribe = viewModel.canTranscribe,
|
||||
isRecording = viewModel.isRecording,
|
||||
messageLog = viewModel.dataLog,
|
||||
onBenchmarkTapped = viewModel::benchmark,
|
||||
onTranscribeSampleTapped = viewModel::transcribeSample,
|
||||
onRecordTapped = viewModel::toggleRecord
|
||||
)
|
||||
@ -32,7 +30,6 @@ private fun MainScreen(
|
||||
canTranscribe: Boolean,
|
||||
isRecording: Boolean,
|
||||
messageLog: String,
|
||||
onBenchmarkTapped: () -> Unit,
|
||||
onTranscribeSampleTapped: () -> Unit,
|
||||
onRecordTapped: () -> Unit
|
||||
) {
|
||||
@ -48,11 +45,8 @@ private fun MainScreen(
|
||||
.padding(innerPadding)
|
||||
.padding(16.dp)
|
||||
) {
|
||||
Column(verticalArrangement = Arrangement.SpaceBetween) {
|
||||
Row(horizontalArrangement = Arrangement.SpaceBetween, modifier = Modifier.fillMaxWidth()) {
|
||||
BenchmarkButton(enabled = canTranscribe, onClick = onBenchmarkTapped)
|
||||
Row(horizontalArrangement = Arrangement.SpaceBetween) {
|
||||
TranscribeSampleButton(enabled = canTranscribe, onClick = onTranscribeSampleTapped)
|
||||
}
|
||||
RecordButton(
|
||||
enabled = canTranscribe,
|
||||
isRecording = isRecording,
|
||||
@ -66,16 +60,7 @@ private fun MainScreen(
|
||||
|
||||
@Composable
|
||||
private fun MessageLog(log: String) {
|
||||
SelectionContainer() {
|
||||
Text(modifier = Modifier.verticalScroll(rememberScrollState()), text = log)
|
||||
}
|
||||
}
|
||||
|
||||
@Composable
|
||||
private fun BenchmarkButton(enabled: Boolean, onClick: () -> Unit) {
|
||||
Button(onClick = onClick, enabled = enabled) {
|
||||
Text("Benchmark")
|
||||
}
|
||||
}
|
||||
|
||||
@Composable
|
||||
|
||||
@ -41,15 +41,10 @@ class MainScreenViewModel(private val application: Application) : ViewModel() {
|
||||
|
||||
init {
|
||||
viewModelScope.launch {
|
||||
printSystemInfo()
|
||||
loadData()
|
||||
}
|
||||
}
|
||||
|
||||
private suspend fun printSystemInfo() {
|
||||
printMessage(String.format("System Info: %s\n", WhisperContext.getSystemInfo()));
|
||||
}
|
||||
|
||||
private suspend fun loadData() {
|
||||
printMessage("Loading data...\n")
|
||||
try {
|
||||
@ -86,29 +81,10 @@ class MainScreenViewModel(private val application: Application) : ViewModel() {
|
||||
//whisperContext = WhisperContext.createContextFromFile(firstModel.absolutePath)
|
||||
}
|
||||
|
||||
fun benchmark() = viewModelScope.launch {
|
||||
runBenchmark(6)
|
||||
}
|
||||
|
||||
fun transcribeSample() = viewModelScope.launch {
|
||||
transcribeAudio(getFirstSample())
|
||||
}
|
||||
|
||||
private suspend fun runBenchmark(nthreads: Int) {
|
||||
if (!canTranscribe) {
|
||||
return
|
||||
}
|
||||
|
||||
canTranscribe = false
|
||||
|
||||
printMessage("Running benchmark. This will take minutes...\n")
|
||||
whisperContext?.benchMemory(nthreads)?.let{ printMessage(it) }
|
||||
printMessage("\n")
|
||||
whisperContext?.benchGgmlMulMat(nthreads)?.let{ printMessage(it) }
|
||||
|
||||
canTranscribe = true
|
||||
}
|
||||
|
||||
private suspend fun getFirstSample(): File = withContext(Dispatchers.IO) {
|
||||
samplesPath.listFiles()!!.first()
|
||||
}
|
||||
@ -138,14 +114,11 @@ class MainScreenViewModel(private val application: Application) : ViewModel() {
|
||||
canTranscribe = false
|
||||
|
||||
try {
|
||||
printMessage("Reading wave samples... ")
|
||||
printMessage("Reading wave samples...\n")
|
||||
val data = readAudioSamples(file)
|
||||
printMessage("${data.size / (16000 / 1000)} ms\n")
|
||||
printMessage("Transcribing data...\n")
|
||||
val start = System.currentTimeMillis()
|
||||
val text = whisperContext?.transcribeData(data)
|
||||
val elapsed = System.currentTimeMillis() - start
|
||||
printMessage("Done ($elapsed ms): $text\n")
|
||||
printMessage("Done: $text\n")
|
||||
} catch (e: Exception) {
|
||||
Log.w(LOG_TAG, e)
|
||||
printMessage("${e.localizedMessage}\n")
|
||||
|
||||
@ -27,14 +27,6 @@ class WhisperContext private constructor(private var ptr: Long) {
|
||||
}
|
||||
}
|
||||
|
||||
suspend fun benchMemory(nthreads: Int): String = withContext(scope.coroutineContext) {
|
||||
return@withContext WhisperLib.benchMemcpy(nthreads)
|
||||
}
|
||||
|
||||
suspend fun benchGgmlMulMat(nthreads: Int): String = withContext(scope.coroutineContext) {
|
||||
return@withContext WhisperLib.benchGgmlMulMat(nthreads)
|
||||
}
|
||||
|
||||
suspend fun release() = withContext(scope.coroutineContext) {
|
||||
if (ptr != 0L) {
|
||||
WhisperLib.freeContext(ptr)
|
||||
@ -74,10 +66,6 @@ class WhisperContext private constructor(private var ptr: Long) {
|
||||
}
|
||||
return WhisperContext(ptr)
|
||||
}
|
||||
|
||||
fun getSystemInfo(): String {
|
||||
return WhisperLib.getSystemInfo()
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
@ -86,7 +74,6 @@ 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()
|
||||
@ -97,24 +84,11 @@ 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: $cpuInfo")
|
||||
if (cpuInfo.contains("fphp")) {
|
||||
Log.d(LOG_TAG, "CPU supports fp16 arithmetic")
|
||||
loadV8fp16 = true
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (loadVfpv4) {
|
||||
Log.d(LOG_TAG, "Loading libwhisper_vfpv4.so")
|
||||
System.loadLibrary("whisper_vfpv4")
|
||||
} else if (loadV8fp16) {
|
||||
Log.d(LOG_TAG, "Loading libwhisper_v8fp16_va.so")
|
||||
System.loadLibrary("whisper_v8fp16_va")
|
||||
} else {
|
||||
Log.d(LOG_TAG, "Loading libwhisper.so")
|
||||
System.loadLibrary("whisper")
|
||||
@ -129,9 +103,6 @@ private class WhisperLib {
|
||||
external fun fullTranscribe(contextPtr: Long, audioData: FloatArray)
|
||||
external fun getTextSegmentCount(contextPtr: Long): Int
|
||||
external fun getTextSegment(contextPtr: Long, index: Int): String
|
||||
external fun getSystemInfo(): String
|
||||
external fun benchMemcpy(nthread: Int): String
|
||||
external fun benchGgmlMulMat(nthread: Int): String
|
||||
}
|
||||
}
|
||||
|
||||
@ -139,10 +110,6 @@ private fun isArmEabiV7a(): Boolean {
|
||||
return Build.SUPPORTED_ABIS[0].equals("armeabi-v7a")
|
||||
}
|
||||
|
||||
private fun isArmEabiV8a(): Boolean {
|
||||
return Build.SUPPORTED_ABIS[0].equals("arm64-v8a")
|
||||
}
|
||||
|
||||
private fun cpuInfo(): String? {
|
||||
return try {
|
||||
File("/proc/cpuinfo").inputStream().bufferedReader().use {
|
||||
|
||||
@ -13,14 +13,3 @@ ifeq ($(TARGET_ARCH_ABI),armeabi-v7a)
|
||||
LOCAL_CFLAGS += -mfpu=neon-vfpv4
|
||||
include $(BUILD_SHARED_LIBRARY)
|
||||
endif
|
||||
|
||||
ifeq ($(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
|
||||
|
||||
|
||||
@ -6,7 +6,6 @@
|
||||
#include <sys/sysinfo.h>
|
||||
#include <string.h>
|
||||
#include "whisper.h"
|
||||
#include "ggml.h"
|
||||
|
||||
#define UNUSED(x) (void)(x)
|
||||
#define TAG "JNI"
|
||||
@ -215,29 +214,3 @@ Java_com_whispercppdemo_whisper_WhisperLib_00024Companion_getTextSegment(
|
||||
jstring string = (*env)->NewStringUTF(env, text);
|
||||
return string;
|
||||
}
|
||||
|
||||
JNIEXPORT jstring JNICALL
|
||||
Java_com_whispercppdemo_whisper_WhisperLib_00024Companion_getSystemInfo(
|
||||
JNIEnv *env, jobject thiz
|
||||
) {
|
||||
UNUSED(thiz);
|
||||
const char *sysinfo = whisper_print_system_info();
|
||||
jstring string = (*env)->NewStringUTF(env, sysinfo);
|
||||
return string;
|
||||
}
|
||||
|
||||
JNIEXPORT jstring JNICALL
|
||||
Java_com_whispercppdemo_whisper_WhisperLib_00024Companion_benchMemcpy(JNIEnv *env, jobject thiz,
|
||||
jint n_threads) {
|
||||
UNUSED(thiz);
|
||||
const char *bench_ggml_memcpy = whisper_bench_memcpy_str(n_threads);
|
||||
jstring string = (*env)->NewStringUTF(env, bench_ggml_memcpy);
|
||||
}
|
||||
|
||||
JNIEXPORT jstring JNICALL
|
||||
Java_com_whispercppdemo_whisper_WhisperLib_00024Companion_benchGgmlMulMat(JNIEnv *env, jobject thiz,
|
||||
jint n_threads) {
|
||||
UNUSED(thiz);
|
||||
const char *bench_ggml_mul_mat = whisper_bench_ggml_mul_mat_str(n_threads);
|
||||
jstring string = (*env)->NewStringUTF(env, bench_ggml_mul_mat);
|
||||
}
|
||||
|
||||
@ -24,5 +24,3 @@ Also, don't forget to add the `-DGGML_USE_ACCELERATE` compiler flag in Build Pha
|
||||
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">
|
||||
|
||||
In this project, it also added `-O3 -DNDEBUG` to `Other C Flags`, but adding flags to app proj is not ideal in real world (applies to all C/C++ files), consider splitting xcodeproj in workspace in your own project.
|
||||
|
||||
@ -296,10 +296,6 @@
|
||||
IPHONEOS_DEPLOYMENT_TARGET = 16.0;
|
||||
MTL_ENABLE_DEBUG_INFO = NO;
|
||||
MTL_FAST_MATH = YES;
|
||||
OTHER_CFLAGS = (
|
||||
"-O3",
|
||||
"-DNDEBUG",
|
||||
);
|
||||
SDKROOT = iphoneos;
|
||||
VALIDATE_PRODUCT = YES;
|
||||
};
|
||||
|
||||
@ -7,9 +7,8 @@ To use:
|
||||
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" [^2] build configuration under "Run", then deploy and run to your device.
|
||||
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.
|
||||
[^2]: The `Release` build can boost performance of transcription. In this project, it also added `-O3 -DNDEBUG` to `Other C Flags`, but adding flags to app proj is not ideal in real world (applies to all C/C++ files), consider splitting xcodeproj in workspace in your own project.
|
||||
|
||||
|
||||
|
||||
@ -430,10 +430,6 @@
|
||||
LLVM_LTO = YES;
|
||||
MACOSX_DEPLOYMENT_TARGET = 13.0;
|
||||
MARKETING_VERSION = 1.0;
|
||||
OTHER_CFLAGS = (
|
||||
"-O3",
|
||||
"-DNDEBUG",
|
||||
);
|
||||
PRODUCT_BUNDLE_IDENTIFIER = com.whispercppdemo.WhisperCppDemo;
|
||||
PRODUCT_NAME = "$(TARGET_NAME)";
|
||||
SDKROOT = auto;
|
||||
|
||||
@ -1,70 +0,0 @@
|
||||
# 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
|
||||
@ -6,7 +6,7 @@ using the [convert-pt-to-ggml.py](convert-pt-to-ggml.py) script. You can either
|
||||
the `ggml` files yourself using the conversion script, or you can use the [download-ggml-model.sh](download-ggml-model.sh)
|
||||
script to download the already converted models. Currently, they are hosted on the following locations:
|
||||
|
||||
- https://huggingface.co/ggerganov/whisper.cpp
|
||||
- https://huggingface.co/datasets/ggerganov/whisper.cpp
|
||||
- https://ggml.ggerganov.com
|
||||
|
||||
Sample usage:
|
||||
@ -23,7 +23,7 @@ You can now use it like this:
|
||||
|
||||
A third option to obtain the model files is to download them from Hugging Face:
|
||||
|
||||
https://huggingface.co/ggerganov/whisper.cpp/tree/main
|
||||
https://huggingface.co/datasets/ggerganov/whisper.cpp/tree/main
|
||||
|
||||
## Available models
|
||||
|
||||
|
||||
@ -79,11 +79,11 @@ dir_model = sys.argv[1]
|
||||
dir_whisper = sys.argv[2]
|
||||
dir_out = sys.argv[3]
|
||||
|
||||
with open(dir_model + "/vocab.json", "r", encoding="utf8") as f:
|
||||
with open(dir_model + "/vocab.json", "r") as f:
|
||||
encoder = json.load(f)
|
||||
with open(dir_model + "/added_tokens.json", "r", encoding="utf8") as f:
|
||||
with open(dir_model + "/added_tokens.json", "r") as f:
|
||||
encoder_added = json.load(f)
|
||||
with open(dir_model + "/config.json", "r", encoding="utf8") as f:
|
||||
with open(dir_model + "/config.json", "r") as f:
|
||||
hparams = json.load(f)
|
||||
|
||||
model = WhisperForConditionalGeneration.from_pretrained(dir_model)
|
||||
|
||||
@ -40,7 +40,7 @@ if exist "ggml-%model%.bin" (
|
||||
goto :eof
|
||||
)
|
||||
|
||||
PowerShell -NoProfile -ExecutionPolicy Bypass -Command "Invoke-WebRequest -Uri https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-%model%.bin -OutFile ggml-%model%.bin"
|
||||
PowerShell -NoProfile -ExecutionPolicy Bypass -Command "Invoke-WebRequest -Uri https://huggingface.co/datasets/ggerganov/whisper.cpp/resolve/main/ggml-%model%.bin -OutFile ggml-%model%.bin"
|
||||
|
||||
if %ERRORLEVEL% neq 0 (
|
||||
echo Failed to download ggml model %model%
|
||||
|
||||
@ -6,7 +6,7 @@
|
||||
#src="https://ggml.ggerganov.com"
|
||||
#pfx="ggml-model-whisper"
|
||||
|
||||
src="https://huggingface.co/ggerganov/whisper.cpp"
|
||||
src="https://huggingface.co/datasets/ggerganov/whisper.cpp"
|
||||
pfx="resolve/main/ggml"
|
||||
|
||||
# get the path of this script
|
||||
|
||||
988
whisper.cpp
988
whisper.cpp
File diff suppressed because it is too large
Load Diff
155
whisper.h
155
whisper.h
@ -66,7 +66,6 @@ extern "C" {
|
||||
//
|
||||
|
||||
struct whisper_context;
|
||||
struct whisper_state;
|
||||
|
||||
typedef int whisper_token;
|
||||
|
||||
@ -102,20 +101,11 @@ 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);
|
||||
|
||||
// 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);
|
||||
// Frees all memory allocated by the model.
|
||||
WHISPER_API void whisper_free(struct whisper_context * ctx);
|
||||
|
||||
// Convert RAW PCM audio to log mel spectrogram.
|
||||
// The resulting spectrogram is stored inside the default state of the provided whisper context.
|
||||
// The resulting spectrogram is stored inside the provided whisper context.
|
||||
// Returns 0 on success
|
||||
WHISPER_API int whisper_pcm_to_mel(
|
||||
struct whisper_context * ctx,
|
||||
@ -123,30 +113,17 @@ extern "C" {
|
||||
int n_samples,
|
||||
int n_threads);
|
||||
|
||||
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.
|
||||
// The resulting spectrogram is stored inside the provided whisper context.
|
||||
// Returns 0 on success
|
||||
WHISPER_API int whisper_pcm_to_mel_phase_vocoder(
|
||||
struct whisper_context * ctx,
|
||||
const float * samples,
|
||||
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 default state of the provided whisper context.
|
||||
// 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
|
||||
// Returns 0 on success
|
||||
@ -156,14 +133,7 @@ extern "C" {
|
||||
int n_len,
|
||||
int n_mel);
|
||||
|
||||
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.
|
||||
// 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.
|
||||
// Returns 0 on success
|
||||
@ -172,12 +142,6 @@ 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.
|
||||
@ -191,14 +155,6 @@ 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
|
||||
@ -234,44 +190,20 @@ extern "C" {
|
||||
int n_threads,
|
||||
float * lang_probs);
|
||||
|
||||
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);
|
||||
|
||||
WHISPER_API int whisper_model_n_vocab (struct whisper_context * ctx);
|
||||
WHISPER_API int whisper_model_n_audio_ctx (struct whisper_context * ctx);
|
||||
WHISPER_API int whisper_model_n_audio_state(struct whisper_context * ctx);
|
||||
WHISPER_API int whisper_model_n_audio_head (struct whisper_context * ctx);
|
||||
WHISPER_API int whisper_model_n_audio_layer(struct whisper_context * ctx);
|
||||
WHISPER_API int whisper_model_n_text_ctx (struct whisper_context * ctx);
|
||||
WHISPER_API int whisper_model_n_text_state (struct whisper_context * ctx);
|
||||
WHISPER_API int whisper_model_n_text_head (struct whisper_context * ctx);
|
||||
WHISPER_API int whisper_model_n_text_layer (struct whisper_context * ctx);
|
||||
WHISPER_API int whisper_model_n_mels (struct whisper_context * ctx);
|
||||
WHISPER_API int whisper_model_f16 (struct whisper_context * ctx);
|
||||
WHISPER_API int whisper_model_type (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_from_state(struct whisper_state * state);
|
||||
WHISPER_API float * whisper_get_logits(struct whisper_context * ctx);
|
||||
|
||||
// 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);
|
||||
WHISPER_API const char * whisper_model_type_readable(struct whisper_context * ctx);
|
||||
|
||||
|
||||
// Special tokens
|
||||
WHISPER_API whisper_token whisper_token_eot (struct whisper_context * ctx);
|
||||
@ -286,7 +218,7 @@ extern "C" {
|
||||
WHISPER_API whisper_token whisper_token_translate (void);
|
||||
WHISPER_API whisper_token whisper_token_transcribe(void);
|
||||
|
||||
// Performance information from the default state.
|
||||
// Performance information
|
||||
WHISPER_API void whisper_print_timings(struct whisper_context * ctx);
|
||||
WHISPER_API void whisper_reset_timings(struct whisper_context * ctx);
|
||||
|
||||
@ -304,19 +236,18 @@ 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, struct whisper_state * state, int n_new, void * user_data);
|
||||
typedef void (*whisper_new_segment_callback)(struct whisper_context * ctx, 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, struct whisper_state * state, void * user_data);
|
||||
typedef bool (*whisper_encoder_begin_callback)(struct whisper_context * ctx, 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,
|
||||
@ -403,7 +334,6 @@ 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,
|
||||
@ -411,16 +341,7 @@ extern "C" {
|
||||
const float * samples,
|
||||
int n_samples);
|
||||
|
||||
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.
|
||||
// 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.
|
||||
WHISPER_API int whisper_full_parallel(
|
||||
@ -430,56 +351,40 @@ 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_from_state(struct whisper_state * state);
|
||||
WHISPER_API int whisper_full_n_segments(struct whisper_context * ctx);
|
||||
|
||||
// Language id associated with the context's default state
|
||||
// Language id associated with the current context
|
||||
WHISPER_API int whisper_full_lang_id(struct whisper_context * ctx);
|
||||
|
||||
// Language id associated with the provided state
|
||||
WHISPER_API int whisper_full_lang_id_from_state(struct whisper_state * state);
|
||||
// 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);
|
||||
|
||||
// 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 the text of the specified segment.
|
||||
WHISPER_API const char * whisper_full_get_segment_text(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 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 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);
|
||||
// Get number of tokens in the specified segment.
|
||||
WHISPER_API int whisper_full_n_tokens(struct whisper_context * ctx, 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);
|
||||
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
|
||||
// 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_from_state(struct whisper_state * state, 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);
|
||||
|
||||
// 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);
|
||||
// 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);
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
// Temporary helpers needed for exposing ggml interface
|
||||
|
||||
WHISPER_API int whisper_bench_memcpy(int n_threads);
|
||||
WHISPER_API const char * whisper_bench_memcpy_str(int n_threads);
|
||||
WHISPER_API int whisper_bench_ggml_mul_mat(int n_threads);
|
||||
WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads);
|
||||
|
||||
#ifdef __cplusplus
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user