whisper_ggml 1.8.0

OpenAI Whisper ASR (Automatic Speech Recognition) for Flutter

Whisper GGML

OpenAI Whisper ASR (Automatic Speech Recognition) for Flutter using Whisper.cpp.

Supported platforms

Platform	Supported
Android	✅
iOS	✅
MacOS	✅

Features

• Automatic Speech Recognition integration for Flutter apps.

• Supports automatic model downloading and initialization. Can be configured to work fully offline by using assets models (see example folder).

• Seamless iOS and Android support with optimized performance.

• Can be configured to use specific language ("en", "fr", "de", etc) or auto-detect ("auto").

• Utilizes CORE ML for enhanced processing on iOS devices.

Installation

To use this library in your Flutter project, follow these steps:

1. Add the library to your Flutter project's pubspec.yaml:

dependencies:
  whisper_ggml: ^1.8.0

2. Run flutter pub get to install the package.

Usage

To integrate Whisper ASR in your Flutter app:

1. Import the package:

import 'package:whisper_ggml/whisper_ggml.dart';

2. Pick your model. Smaller models are more performant, but the accuracy may be lower. Recommended models are tiny and small.

final model = WhisperModel.tiny;

3. Declare WhisperController and use it for transcription:

final controller = WhisperController();

final result = await controller.transcribe(
    model: model, /// Selected WhisperModel
    audioPath: audioPath, /// Path to .wav file
    lang: 'en', /// Language to transcribe
    initialPrompt: 'Optional text to bias decoding', /// See note below
);

The optional initialPrompt is passed to whisper.cpp as whisper_full_params.initial_prompt. Whisper uses it to bias decoding toward the vocabulary, proper nouns, and punctuation it contains — useful for domain-specific transcription (medical, legal, product names, etc.) where those words otherwise get misrecognised. Leave it null (the default) to disable biasing. Note that decoding also mimics the prompt's style, so an unpunctuated prompt tends to produce unpunctuated output.

The optional noContext flag sets whisper_full_params.no_context (equivalent to Python whisper's condition_on_previous_text=False). When true, whisper.cpp does not feed prior-segment transcripts into the decoder as context — useful for short, independent utterances where carry-over context can cause hallucinated repetition. Defaults to false, matching whisper.cpp's default.

4. Use the result variable to access the transcription result:

if (result?.transcription.text != null) {
    /// Do something with the transcription
    print(result!.transcription.text);
}

Notes

Transcription processing time is about 5x times faster when running in release mode.