Movatterモバイル変換

[0]ホーム
URL:

Skip to content

Navigation Menu

Sign in
Appearance settings

Search code, repositories, users, issues, pull requests...

Provide feedback

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly

 Sign up
Appearance settings

Port of OpenAI's Whisper model in C/C++

License

NotificationsYou must be signed in to change notification settings

macocha/whisper.cpp

 
 

Repository files navigation

Actions StatusLicense: MITnpm

Stable:v1.2.1 /Roadmap | F.A.Q.

High-performance inference ofOpenAI's Whisper automatic speech recognition (ASR) model:

  • Plain C/C++ implementation without dependencies
  • Apple silicon first-class citizen - optimized via Arm Neon and Accelerate framework
  • AVX intrinsics support for x86 architectures
  • VSX intrinsics support for POWER architectures
  • Mixed F16 / F32 precision
  • Low memory usage (Flash Attention)
  • Zero memory allocations at runtime
  • Runs on the CPU
  • C-style API

Supported platforms:

The entire implementation of the model is contained in 2 source files:

Having such a lightweight implementation of the model allows to easily integrate it in different platforms and applications.As an example, here is a video of running the model on an iPhone 13 device - fully offline, on-device:whisper.objc

whisper-iphone-13-mini-2.mp4

You can also easily make your own offline voice assistant application:command

command-0.mp4

Or you can even run it straight in the browser:talk.wasm

Implementation details

  • The core tensor operations are implemented in C (ggml.h /ggml.c)
  • The transformer model and the high-level C-style API are implemented in C++ (whisper.h /whisper.cpp)
  • Sample usage is demonstrated inmain.cpp
  • Sample real-time audio transcription from the microphone is demonstrated instream.cpp
  • Various other examples are available in theexamples folder

The tensor operators are optimized heavily for Apple silicon CPUs. Depending on the computation size, Arm Neon SIMDinstrisics or CBLAS Accelerate framework routines are used. The latter are especially effective for bigger sizes sincethe Accelerate framework utilizes the special-purpose AMX coprocessor available in modern Apple products.

Quick start

First, download one of the Whisper models converted inggml format. For example:

bash ./models/download-ggml-model.sh base.en

Now build themain example and transcribe an audio file like this:

# build the main examplemake# transcribe an audio file./main -f samples/jfk.wav

For a quick demo, simply runmake base.en:

$makebase.encc  -I.              -O3 -std=c11   -pthread -DGGML_USE_ACCELERATE   -cggml.c -oggml.oc++ -I. -I./examples -O3 -std=c++11 -pthread -cwhisper.cpp -owhisper.oc++ -I. -I./examples -O3 -std=c++11 -pthreadexamples/main/main.cppwhisper.oggml.o -omain  -frameworkAccelerate./main -husage: ./main [options]file0.wavfile1.wav ...options:  -h,        --help              [default]showthishelpmessageandexit  -tN,      --threadsN         [4      ]numberofthreadstouseduringcomputation  -pN,      --processorsN      [1      ]numberofprocessorstouseduringcomputation  -otN,     --offset-tN        [0      ]timeoffsetinmilliseconds  -onN,     --offset-nN        [0      ]segmentindexoffset  -dN,     --durationN        [0      ]durationofaudiotoprocessinmilliseconds  -mcN,     --max-contextN     [-1     ]maximumnumberoftextcontexttokenstostore  -mlN,     --max-lenN         [0      ]maximumsegmentlengthincharacters  -boN,     --best-ofN         [5      ]numberofbestcandidatestokeep  -bsN,     --beam-sizeN       [-1     ]beamsizeforbeamsearch  -wtN,     --word-tholdN      [0.01   ]wordtimestampprobabilitythreshold  -etN,     --entropy-tholdN   [2.40   ]entropythresholdfordecoderfail  -lptN,    --logprob-tholdN   [-1.00  ]logprobabilitythresholdfordecoderfail  -su,       --speed-up          [false  ]speedupaudiobyx2 (reducedaccuracy)  -tr,       --translate         [false  ]translatefromsourcelanguagetoenglish  -di,       --diarize           [false  ]stereoaudiodiarization  -nf,       --no-fallback       [false  ]donotusetemperaturefallbackwhiledecoding  -otxt,     --output-txt        [false  ]outputresultinatextfile  -ovtt,     --output-vtt        [false  ]outputresultinavttfile  -osrt,     --output-srt        [false  ]outputresultinasrtfile  -owts,     --output-words      [false  ]outputscriptforgeneratingkaraokevideo  -ocsv,     --output-csv        [false  ]outputresultinaCSVfile  -ofFNAME, --output-fileFNAME [       ]outputfilepath (withoutfileextension)  -ps,       --print-special     [false  ]printspecialtokens  -pc,       --print-colors      [false  ]printcolors  -pp,       --print-progress    [false  ]printprogress  -nt,       --no-timestamps     [true   ]donotprinttimestamps  -lLANG,   --languageLANG     [en     ]spokenlanguage ('auto'forauto-detect)             --promptPROMPT     [       ]initialprompt  -mFNAME,  --modelFNAME       [models/ggml-base.en.bin]modelpath  -fFNAME,  --fileFNAME        [       ]inputWAVfilepathbash ./models/download-ggml-model.shbase.enDownloadingggmlmodelbase.en ...ggml-base.en.bin100%[========================>]141.11M6.34MB/sin24sDone!Model'base.en'savedin'models/ggml-base.en.bin'Youcannowuseitlikethis:$ ./main -mmodels/ggml-base.en.bin -fsamples/jfk.wav===============================================Runningbase.enonallsamplesin ./samples ...===============================================----------------------------------------------[+]Runningbase.enonsamples/jfk.wav ... (run'ffplay samples/jfk.wav'tolisten)----------------------------------------------whisper_init_from_file:loadingmodelfrom'models/ggml-base.en.bin'whisper_model_load:loadingmodelwhisper_model_load:n_vocab       =51864whisper_model_load:n_audio_ctx   =1500whisper_model_load:n_audio_state =512whisper_model_load:n_audio_head  =8whisper_model_load:n_audio_layer =6whisper_model_load:n_text_ctx    =448whisper_model_load:n_text_state  =512whisper_model_load:n_text_head   =8whisper_model_load:n_text_layer  =6whisper_model_load:n_mels        =80whisper_model_load:f16           =1whisper_model_load:type          =2whisper_model_load:memrequired  =215.00MB (+6.00MBperdecoder)whisper_model_load:kvselfsize  =5.25MBwhisper_model_load:kvcrosssize =17.58MBwhisper_model_load:adding1607extratokenswhisper_model_load:modelctx     =140.60MBwhisper_model_load:modelsize    =140.54MBsystem_info:n_threads =4 /10 |AVX =0 |AVX2 =0 |AVX512 =0 |FMA =0 |NEON =1 |ARM_FMA =1 |F16C =0 |FP16_VA =1 |WASM_SIMD =0 |BLAS =1 |SSE3 =0 |VSX =0 |main:processing'samples/jfk.wav' (176000samples,11.0sec),4threads,1processors,lang =en,task =transcribe,timestamps =1 ...[00:00:00.000 -->00:00:11.000]AndsomyfellowAmericans,asknotwhatyourcountrycandoforyou,askwhatyoucandoforyourcountry.whisper_print_timings:fallbacks =0p /0hwhisper_print_timings:loadtime =113.81mswhisper_print_timings:meltime =15.40mswhisper_print_timings:sampletime =11.58ms /27runs (0.43msperrun)whisper_print_timings:encodetime =266.60ms /1runs (266.60msperrun)whisper_print_timings:decodetime =66.11ms /27runs (2.45msperrun)whisper_print_timings:totaltime =476.31ms

The command downloads thebase.en model converted to customggml format and runs the inference on all.wav samples in the foldersamples.

For detailed usage instructions, run:./main -h

Note that themain example currently runs only with 16-bit WAV files, so make sure to convert your input before running the tool.For example, you can useffmpeg like this:

ffmpeg -iinput.mp3 -ar16000 -ac1 -c:apcm_s16leoutput.wav

More audio samples

If you want some extra audio samples to play with, simply run:

make samples

This will download a few more audio files from Wikipedia and convert them to 16-bit WAV format viaffmpeg.

You can download and run the other models as follows:

make tiny.enmake tinymake base.enmake basemake small.enmake smallmake medium.enmake mediummake large-v1make large

Memory usage

ModelDiskMemSHA
tiny75 MB~125 MBbd577a113a864445d4c299885e0cb97d4ba92b5f
base142 MB~210 MB465707469ff3a37a2b9b8d8f89f2f99de7299dac
small466 MB~600 MB55356645c2b361a969dfd0ef2c5a50d530afd8d5
medium1.5 GB~1.7 GBfd9727b6e1217c2f614f9b698455c4ffd82463b4
large2.9 GB~3.3 GB0f4c8e34f21cf1a914c59d8b3ce882345ad349d6

Limitations

  • Inference only
  • No GPU support (yet)

Another example

Here is another example of transcribing a3:24 min speechin about half a minute on a MacBook M1 Pro, usingmedium.en model:

Expand to see the result
$ ./main -mmodels/ggml-medium.en.bin -fsamples/gb1.wav -t8whisper_init_from_file:loadingmodelfrom'models/ggml-medium.en.bin'whisper_model_load:loadingmodelwhisper_model_load:n_vocab       =51864whisper_model_load:n_audio_ctx   =1500whisper_model_load:n_audio_state =1024whisper_model_load:n_audio_head  =16whisper_model_load:n_audio_layer =24whisper_model_load:n_text_ctx    =448whisper_model_load:n_text_state  =1024whisper_model_load:n_text_head   =16whisper_model_load:n_text_layer  =24whisper_model_load:n_mels        =80whisper_model_load:f16           =1whisper_model_load:type          =4whisper_model_load:memrequired  =1720.00MB (+43.00MBperdecoder)whisper_model_load:kvselfsize  =42.00MBwhisper_model_load:kvcrosssize =140.62MBwhisper_model_load:adding1607extratokenswhisper_model_load:modelctx     =1462.35MBwhisper_model_load:modelsize    =1462.12MBsystem_info:n_threads =8 /10 |AVX =0 |AVX2 =0 |AVX512 =0 |FMA =0 |NEON =1 |ARM_FMA =1 |F16C =0 |FP16_VA =1 |WASM_SIMD =0 |BLAS =1 |SSE3 =0 |VSX =0 |main:processing'samples/gb1.wav' (3179750samples,198.7sec),8threads,1processors,lang =en,task =transcribe,timestamps =1 ...[00:00:00.000 -->00:00:08.000]MyfellowAmericans,thisdayhasbroughtterriblenewsandgreatsadnesstoourcountry.[00:00:08.000 -->00:00:17.000]Atnineo'clock this morning, Mission Control in Houston lost contact with our Space Shuttle Columbia.[00:00:17.000 -->00:00:23.000]Ashorttimelater,debriswasseenfallingfromtheskiesaboveTexas.[00:00:23.000 -->00:00:29.000]TheColumbia's lost. There are no survivors.[00:00:29.000 -->00:00:32.000]Onboardwasacrewofseven.[00:00:32.000 -->00:00:39.000]ColonelRickHusband,LieutenantColonelMichaelAnderson,CommanderLaurelClark,[00:00:39.000 -->00:00:48.000]CaptainDavidBrown,CommanderWilliamMcCool,Dr.KultnaShavla,andIlanRamon,[00:00:48.000 -->00:00:52.000]acolonelintheIsraeliAirForce.[00:00:52.000 -->00:00:58.000]Thesemenandwomenassumedgreatriskintheservicetoallhumanity.[00:00:58.000 -->00:01:03.000]Inanagewhenspaceflighthascometoseemalmostroutine,[00:01:03.000 -->00:01:07.000]itiseasytooverlookthedangersoftravelbyrocket[00:01:07.000 -->00:01:12.000]andthedifficultiesofnavigatingthefierceouteratmosphereoftheEarth.[00:01:12.000 -->00:01:18.000]Theseastronautsknewthedangers,andtheyfacedthemwillingly,[00:01:18.000 -->00:01:23.000]knowingtheyhadahighandnoblepurposeinlife.[00:01:23.000 -->00:01:31.000]Becauseoftheircourageanddaringandidealism,wewillmissthemallthemore.[00:01:31.000 -->00:01:36.000]AllAmericanstodayarethinkingaswellofthefamiliesofthesemenandwomen[00:01:36.000 -->00:01:40.000]whohavebeengiventhissuddenshockandgrief.[00:01:40.000 -->00:01:45.000]You're not alone. Our entire nation grieves with you,[00:01:45.000 -->00:01:52.000]andthoseyoulovewillalwayshavetherespectandgratitudeofthiscountry.[00:01:52.000 -->00:01:56.000]Thecauseinwhichtheydiedwillcontinue.[00:01:56.000 -->00:02:04.000]Mankindisledintothedarknessbeyondourworldbytheinspirationofdiscovery[00:02:04.000 -->00:02:11.000]andthelongingtounderstand.Ourjourneyintospacewillgoon.[00:02:11.000 -->00:02:16.000]Intheskiestoday,wesawdestructionandtragedy.[00:02:16.000 -->00:02:22.000]Yetfartherthanwecansee,thereiscomfortandhope.[00:02:22.000 -->00:02:29.000]InthewordsoftheprophetIsaiah,"Lift your eyes and look to the heavens[00:02:29.000 --> 00:02:35.000]   who created all these. He who brings out the starry hosts one by one[00:02:35.000 --> 00:02:39.000]   and calls them each by name."[00:02:39.000 -->00:02:46.000]BecauseofHisgreatpowerandmightystrength,notoneofthemismissing.[00:02:46.000 -->00:02:55.000]ThesameCreatorwhonamesthestarsalsoknowsthenamesofthesevensoulswemourntoday.[00:02:55.000 -->00:03:01.000]ThecrewoftheshuttleColumbiadidnotreturnsafelytoearth,[00:03:01.000 -->00:03:05.000]yetwecanpraythatallaresafelyhome.[00:03:05.000 -->00:03:13.000]MayGodblessthegrievingfamilies,andmayGodcontinuetoblessAmerica.[00:03:13.000 -->00:03:19.000]   [Silence]whisper_print_timings:fallbacks =1p /0hwhisper_print_timings:loadtime =569.03mswhisper_print_timings:meltime =146.85mswhisper_print_timings:sampletime =238.66ms /553runs (0.43msperrun)whisper_print_timings:encodetime =18665.10ms /9runs (2073.90msperrun)whisper_print_timings:decodetime =13090.93ms /549runs (23.85msperrun)whisper_print_timings:totaltime =32733.52ms

Real-time audio input example

This is a naive example of performing real-time inference on audio from your microphone.Thestream tool samples the audio every half a second and runs the transcription continously.More info is available inissue #10.

makestream./stream -m ./models/ggml-base.en.bin -t8 --step500 --length5000
rt_esl_csgo_2.mp4

Confidence color-coding

Adding the--print-colors argument will print the transcribed text using an experimental color coding strategyto highlight words with high or low confidence:

image

Controlling the length of the generated text segments (experimental)

For example, to limit the line length to a maximum of 16 characters, simply add-ml 16:

./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -ml16whisper_model_load:loadingmodelfrom'./models/ggml-base.en.bin'...system_info:n_threads =4 /10 |AVX2 =0 |AVX512 =0 |NEON =1 |FP16_VA =1 |WASM_SIMD =0 |BLAS =1 |main:processing'./samples/jfk.wav' (176000samples,11.0sec),4threads,1processors,lang =en,task =transcribe,timestamps =1 ...[00:00:00.000 -->00:00:00.850]Andsomy[00:00:00.850 -->00:00:01.590]fellow[00:00:01.590 -->00:00:04.140]Americans,ask[00:00:04.140 -->00:00:05.660]notwhatyour[00:00:05.660 -->00:00:06.840]countrycando[00:00:06.840 -->00:00:08.430]foryou,ask[00:00:08.430 -->00:00:09.440]whatyoucando[00:00:09.440 -->00:00:10.020]foryour[00:00:10.020 -->00:00:11.000]country.

Word-level timestamp

The--max-len argument can be used to obtain word-level timestamps. Simply use-ml 1:

./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -ml1whisper_model_load:loadingmodelfrom'./models/ggml-base.en.bin'...system_info:n_threads =4 /10 |AVX2 =0 |AVX512 =0 |NEON =1 |FP16_VA =1 |WASM_SIMD =0 |BLAS =1 |main:processing'./samples/jfk.wav' (176000samples,11.0sec),4threads,1processors,lang =en,task =transcribe,timestamps =1 ...[00:00:00.000 -->00:00:00.320]  [00:00:00.320 -->00:00:00.370]And[00:00:00.370 -->00:00:00.690]so[00:00:00.690 -->00:00:00.850]my[00:00:00.850 -->00:00:01.590]fellow[00:00:01.590 -->00:00:02.850]Americans[00:00:02.850 -->00:00:03.300]  ,[00:00:03.300 -->00:00:04.140]ask[00:00:04.140 -->00:00:04.990]not[00:00:04.990 -->00:00:05.410]what[00:00:05.410 -->00:00:05.660]your[00:00:05.660 -->00:00:06.260]country[00:00:06.260 -->00:00:06.600]can[00:00:06.600 -->00:00:06.840]do[00:00:06.840 -->00:00:07.010]for[00:00:07.010 -->00:00:08.170]you[00:00:08.170 -->00:00:08.190]  ,[00:00:08.190 -->00:00:08.430]ask[00:00:08.430 -->00:00:08.910]what[00:00:08.910 -->00:00:09.040]you[00:00:09.040 -->00:00:09.320]can[00:00:09.320 -->00:00:09.440]do[00:00:09.440 -->00:00:09.760]for[00:00:09.760 -->00:00:10.020]your[00:00:10.020 -->00:00:10.510]country[00:00:10.510 -->00:00:11.000]  .

Karaoke-style movie generation (experimental)

Themain example provides support for output of karaoke-style movies, where thecurrently pronounced word is highlighted. Use the-wts argument and run the generated bash script.This requires to haveffmpeg installed.

Here are a few"typical" examples:

./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -owtssource ./samples/jfk.wav.wtsffplay ./samples/jfk.wav.mp4
jfk.wav.mp4
./main -m ./models/ggml-base.en.bin -f ./samples/mm0.wav -owtssource ./samples/mm0.wav.wtsffplay ./samples/mm0.wav.mp4
mm0.wav.mp4
./main -m ./models/ggml-base.en.bin -f ./samples/gb0.wav -owtssource ./samples/gb0.wav.wtsffplay ./samples/gb0.wav.mp4
gb0.wav.mp4

Video comparison of different models

Use theextra/bench-wts.sh script to generate a video in the following format:

./extra/bench-wts.shsamples/jfk.wavffplay ./samples/jfk.wav.all.mp4
jfk.wav.all.mp4

Benchmarks

In order to have an objective comparison of the performance of the inference across different system configurations,use thebench tool. The tool simply runs the Encoder part of the model and prints how much time ittook to execute it. The results are summarized in the following Github issue:

Benchmark results

ggml format

The original models are converted to a custom binary format. This allows to pack everything needed into a single file:

  • model parameters
  • mel filters
  • vocabulary
  • weights

You can download the converted models using themodels/download-ggml-model.sh scriptor manually from here:

For more details, see the conversion scriptmodels/convert-pt-to-ggml.py or the READMEinmodels.

Examples

There are various examples of using the library for different projects in theexamples folder.Some of the examples are even ported to run in the browser using WebAssembly. Check them out!

ExampleWebDescription
mainwhisper.wasmTool for translating and transcribing audio using Whisper
benchbench.wasmBenchmark the performance of Whisper on your machine
streamstream.wasmReal-time transcription of raw microphone capture
commandcommand.wasmBasic voice assistant example for receiving voice commands from the mic
talktalk.wasmTalk with a GPT-2 bot
whisper.objciOS mobile application using whisper.cpp
whisper.swiftuiSwiftUI iOS / macOS application using whisper.cpp
whisper.androidAndroid mobile application using whisper.cpp
whisper.nvimSpeech-to-text plugin for Neovim
generate-karaoke.shHelper script to easilygenerate a karaoke video of raw audio capture
livestream.shLivestream audio transcription
yt-wsp.shDownload + transcribe and/or translate any VOD(original)

If you have any kind of feedback about this project feel free to use the Discussions section and open a new topic.You can use theShow and tell categoryto share your own projects that usewhisper.cpp. If you have a question, make sure to check theFrequently asked questions (#126) discussion.

About

Port of OpenAI's Whisper model in C/C++

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • C50.2%
  • C++31.6%
  • Go8.4%
  • Python2.8%
  • CMake2.5%
  • Shell1.8%
  • Other2.7%
[8]ページ先頭
©2009-2025 Movatter.jp