https://github.com/kernelinterrupt/whisper4dart
whisper4dart is a dart wrapper for whisper.cpp, designed to offer an all-in-one speech recognition experience. It can handle most audio file inputs, not just wav.
https://github.com/kernelinterrupt/whisper4dart
cpp dart deep-learning flutter speech-recognition speech-to-text whisper whisper-cpp
Last synced: 2 months ago
JSON representation
whisper4dart is a dart wrapper for whisper.cpp, designed to offer an all-in-one speech recognition experience. It can handle most audio file inputs, not just wav.
- Host: GitHub
- URL: https://github.com/kernelinterrupt/whisper4dart
- Owner: KernelInterrupt
- License: apache-2.0
- Created: 2025-03-03T18:21:53.000Z (about 1 year ago)
- Default Branch: master
- Last Pushed: 2025-07-19T18:24:06.000Z (9 months ago)
- Last Synced: 2025-07-19T21:45:09.603Z (9 months ago)
- Topics: cpp, dart, deep-learning, flutter, speech-recognition, speech-to-text, whisper, whisper-cpp
- Language: Dart
- Homepage:
- Size: 20.4 MB
- Stars: 8
- Watchers: 1
- Forks: 4
- Open Issues: 8
-
Metadata Files:
- Readme: README.md
- Changelog: CHANGELOG.md
- License: LICENSE
Awesome Lists containing this project
README
# whisper4dart
whisper4dart is a dart wrapper for [whisper.cpp](https://github.com/ggerganov/whisper.cpp), designed to offer an all-in-one speech recognition experience. With the built-in decoder/demuxer from ffmpeg, it can handle **most audio/video file** inputs, not just wav.
| Platform | Status |
| :------: | :----: |
| Windows | ✅ |
| Linux | ✅ |
| Android | ✅ |
| iOS | ❌ |
| MacOS | ❌ |
iOS and MacOS version of whisper4dart will be available in the near future. However, we have no intention to support web platform, at least now.
## Getting Started
```
flutter pub add whisper4dart
```
or add following line to your `pubspec.yaml`:
```
whisper4dart:^0.1.4
```
After that,run following command in your terminal:
```
dart rum libmpv_dart:setup --platform
```
(For example,you need to run:`dart run libmpv_dart:setup --platform windows` if you want to setup for windows.)
And then,run:
```
dart run whisper4dart:setup --prebuilt
```
Attention:If you want to build whisper.cpp by yourself instead of using prebuilt libs,run following command:
```
dart run whisper4dart:setup --source
```
OK,now you are ready to use the package,enjoy it!
## How to use
```dart
import 'package:whisper4dart/whisper4dart.dart' as whisper;
final Directory tempDirectory = await getTemporaryDirectory();
final ByteData documentBytes = await rootBundle.load(inputPath);
await File(inputPath).writeAsBytes(
documentBytes.buffer.asUint8List(),
);
//preprocess the file,if the file is not in assets/ ,you don't need to use the code above.
final String logPath = '${tempDirectory.path}/log.txt';
var cparams=whisper4dart.createContextDefaultParams();
//create default parameters,you can modify it on your demand.
var buffer=await rootBundle.load("assets/ggml-base.en.bin");
Uint8List model=buffer.buffer.asUint8List();
//if your model file is not in assets/ ,you dont need to do so,
//and you just need to pass the file path of model to initialize whisper.
//Like this: var model="path/to/your/model";
var whisperModel=whisper.Whisper(model,cparams,outputMode:"plaintext");
//initialize whisper model
//The "outputMode" variable determines the output format. There are four options:
//"plaintext": Outputs plain text
//"txt": Outputs text-formatted strings
//"json": Outputs JSON-formatted strings
//"srt": Outputs SRT-subtitle-formatted strings
String output=await whisperModel.infer(inputPath,
logPath: logPath,
numProcessors: 1,
translate:False,
initialPrompt:"",
startTime:0,
endTime:-1,
useOriginalTime:true);
//The core function whisper.infer takes "inputPath" as the audio file path (e.g., /tmp/jfk.mp3).
//Specifying "logPath" directs whisper4dart to save encoder/demuxer logs in that directory.
//"translate" determines if the output should be translated into English.
//Use "initialPrompt" to set the model's initial prompt.
//"startTime" and "endTime" define the segment of the audio to process (unit: milliseconds).
//Setting "endTime" to -1 means no end cropping is needed.
//When "useOriginalTime" is set to true, the timestamps are based on the start of the original audio,
//not the start of the cropped audio. This means that regardless of whether the audio has
//been cropped, the timestamps always reference the timeline of the original audio.
```
Sample output strings of the four output modes:(input file:jfk.wav)
`plaintext`:
```
And so my fellow Americans, ask not what your country can do for you, ask what you can do for your country.
```
`txt`:
```
And so my fellow Americans, ask not what your country can do for you, ask what you can do for your country.
```
`json`:
```json
[{"multilingual":"false"},{"language":"en"},{"from":"00:00:00.000","to":"00:00:11.000","text":" And so my fellow Americans, ask not what your country can do for you, ask what you can do for your country."}]
```
`srt`:
```
1
00:00:00,000 --> 00:00:11,000
And so my fellow Americans, ask not what your country can do for you, ask what you can do for your country.
```
## Run in isolate
Just use `.inferIsolate()` to replace `.infer()` .
## Output the transcription result in real time
Just use `.inferStream()` to replace `.infer()` .
This method returns a tuple `(ValueNotifier, ValueNotifier)`.(known as a **record** in Dart) You can use these returned notifiers to build your widgets.
The first notifier `(ValueNotifier)` provides the output strings, while the second `(ValueNotifier)` tracks the progress.
(e.g., if the transcription is 50% complete, the value of the progress notifier will be 50.)
> [!CAUTION]
> In this mode,json output is not supported and you have to set numProcessors to 1.
## Acknowledgement
This project leverages contributions from[ libmpv_dart ](https://github.com/Playboy-Player/libmpv_dart)and [whisper.cpp](https://github.com/ggerganov/whisper.cpp)