whisper_edge 0.1.1

On-device speech-to-text for Flutter, powered by whisper.cpp. Offline transcription with timestamps, language detection and model downloads, on iOS and Android.

whisper_edge

On-device speech-to-text for Flutter, powered by whisper.cpp. Fully offline transcription with timestamps, language detection, translation-to-English and built-in model downloads — on iOS and Android.

• iOS: prebuilt whisper.xcframework with Metal + Accelerate acceleration (iOS 16.4+)
• Android: whisper.cpp compiled from source by the NDK (minSdk 24), ARM NEON optimized
• Heavy work runs in a worker isolate — the UI never blocks
• Progress reporting and cancellation
• WAV decoding helpers (any common encoding/sample rate → 16 kHz mono PCM)

Quick start

import 'package:path_provider/path_provider.dart';
import 'package:whisper_edge/whisper_edge.dart';

// 1. Download a model once (≈60 MB for base-q5_1).
final dir = await getApplicationSupportDirectory();
final modelPath = await WhisperModelDownloader().download(
  WhisperModel.baseQ5_1,
  dir,
  onProgress: (received, total) => print('$received / $total'),
);

// 2. Load it.
final transcriber = await WhisperTranscriber.load(modelPath);

// 3. Transcribe a WAV file (or raw 16 kHz mono Float32List PCM).
final result = await transcriber.transcribeWavFile(
  '/path/to/audio.wav',
  language: 'auto',            // or "en", "tr", ...
  onProgress: (p) => print('$p%'),
);

print(result.language);        // detected language
for (final s in result.segments) {
  print('[${s.start} → ${s.end}] ${s.text}');
}

// 4. Free the native model when done.
transcriber.dispose();

transcriber.cancel() aborts an in-flight run. WhisperTranscriber.systemInfo reports the compiled ggml backends.

Choosing a model

Model	Size	Notes
tinyQ5_1	32 MB	fastest, rough quality
baseQ5_1	60 MB	good default for phones
smallQ5_1	190 MB	best quality/speed balance on recent devices
largeV3TurboQ5_0	574 MB	near large-v3 quality, high-end devices

.en variants (tinyEn, baseEn, smallEn) are English-only but more accurate at the same size. Models come from the official ggerganov/whisper.cpp Hugging Face repo; you can also load any compatible ggml .bin file by passing its path straight to WhisperTranscriber.load.

Audio input

Whisper expects 16 kHz mono float32 PCM:

• transcribeWavFile(path) accepts WAV in PCM 8/16/24/32-bit or float32 at any sample rate/channel count and converts automatically.
• transcribe(Float32List pcm) takes raw samples — pair it with a recorder configured for 16 kHz mono (e.g. the record package with AudioEncoder.wav, sampleRate: 16000, numChannels: 1).
• pcm16ToFloat32 / resamplePcm helpers cover live-stream use cases.

For compressed formats (m4a, mp3, ...), convert to WAV first (e.g. with ffmpeg_kit_flutter) — kept out of this package to stay lightweight.

How the native side works

• src/whisper.cpp/ is a pruned, vendored copy of whisper.cpp (see src/whisper.cpp/VENDORED_VERSION), refreshed with tool/fetch_whisper_cpp.sh.
• Android compiles it from source via the plugin's CMake build.
• iOS links the prebuilt ios/Frameworks/whisper.xcframework, rebuilt with tool/build_ios_xcframework.sh (uses whisper.cpp's official build-xcframework.sh).
• Dart talks to a small C shim (src/whisper_edge.c) over FFI, so whisper.cpp upgrades never change the Dart-visible ABI.

License

MIT. whisper.cpp is MIT-licensed by Georgi Gerganov and contributors; the vendored copy retains its license file.