JJ_MKVMaker/packages/whisper_ggml/lib/src/whisper_controller.dart

import 'package:flutter/foundation.dart';
import 'package:path_provider/path_provider.dart';
import 'package:universal_io/io.dart';
import 'package:whisper_ggml/src/models/whisper_model.dart';

import 'models/whisper_result.dart';
import 'whisper.dart';
import 'whisper_live.dart';

class WhisperController {
  String _modelPath = '';
  String? _dir;

  Future<void> initModel(WhisperModel model) async {
    _dir ??= await getModelDir();
    _modelPath = '$_dir/ggml-${model.modelName}.bin';
  }

  /// Start a live transcription session.
  ///
  /// Provide exactly one of [model] and [modelPath]: [model] resolves to the
  /// standard download location (see [getPath]), while [modelPath] runs the
  /// session against any local ggml model file.
  ///
  /// [pcm16Stream] must produce 16 kHz mono little-endian PCM16 audio (e.g.
  /// the `record` package's `startStream` with
  /// `RecordConfig(encoder: AudioEncoder.pcm16bits, sampleRate: 16000,
  /// numChannels: 1)`). Partial transcripts are emitted on
  /// [WhisperLiveSession.partials] while audio flows; when [pcm16Stream]
  /// closes (or [WhisperLiveSession.stop] is called) the session finalizes
  /// and `stop()` returns the full transcript.
  ///
  /// Unlike [transcribe], the model is loaded once for the whole session.
  /// The `gate*` parameters tune the native energy gate that keeps silence
  /// away from the decoder: a chunk counts as voiced when its RMS exceeds
  /// `max(gateVoiceRatio * noiseFloor, gateRmsMin)`, where the adaptive
  /// noise floor is capped at [gateNoiseFloorCap]. Raise the cap for loud
  /// environments; lower [gateRmsMin] for very quiet speakers.
  ///
  /// [keepModelLoaded] parks the model in native memory when the session
  /// stops, so the next session (or one-shot [transcribe]) with the same
  /// model skips the multi-second load — see
  /// [TranscribeRequest.keepModelLoaded]. Release it with [releaseModel].
  Future<WhisperLiveSession> transcribeLive({
    WhisperModel? model,
    String? modelPath,
    required Stream<Uint8List> pcm16Stream,
    String lang = 'en',
    String? initialPrompt,
    bool suppressNonSpeechTokens = false,
    bool keepModelLoaded = false,
    double gateRmsMin = 0.0015,
    double gateVoiceRatio = 2.5,
    double gateNoiseFloorCap = 0.01,
  }) async {
    if ((model == null) == (modelPath == null)) {
      throw ArgumentError('Provide exactly one of model or modelPath.');
    }
    if (model != null) {
      await initModel(model);
    }

    final WhisperLiveSession session = await startWhisperLiveSession(
      modelPath: modelPath ?? _modelPath,
      lang: lang,
      initialPrompt: initialPrompt,
      suppressNonSpeechTokens: suppressNonSpeechTokens,
      keepModelLoaded: keepModelLoaded,
      gateRmsMin: gateRmsMin,
      gateVoiceRatio: gateVoiceRatio,
      gateNoiseFloorCap: gateNoiseFloorCap,
    );

    pcm16Stream.listen(
      session.feed,
      onDone: session.stop,
      onError: (Object e) {
        debugPrint('transcribeLive: audio stream error: $e');
        session.stop();
      },
    );

    return session;
  }

  /// Transcribe [audioPath] with [model].
  ///
  /// When [withSegments] is `true`, the result's
  /// [WhisperTranscribeResponse.segments] carries per-segment timestamps
  /// (`fromTs`/`toTs`) alongside the text; combine with [splitOnWord] to
  /// get one segment per word instead of per phrase.
  ///
  /// [diarize] enables tinydiarize speaker-turn detection: with a `-tdrz`
  /// model ([WhisperModel.smallEnTdrz], English only) and [withSegments],
  /// segments where the speaker changes afterwards have
  /// `speakerTurnNext == true`. Has no effect with regular models.
  ///
  /// [onProgress] reports transcription progress as a 0–100 percentage
  /// (coarse steps) while inference runs.
  ///
  /// [keepModelLoaded] keeps the model resident in native memory after the
  /// request, so the next transcription with the same model skips the
  /// multi-second load — see [TranscribeRequest.keepModelLoaded]. Release
  /// it with [releaseModel].
  Future<TranscribeResult?> transcribe({
    required WhisperModel model,
    required String audioPath,
    String lang = 'en',
    bool diarize = false,
    String? initialPrompt,
    bool noContext = false,
    bool suppressNonSpeechTokens = false,
    bool withSegments = false,
    bool splitOnWord = false,
    bool keepModelLoaded = false,
    void Function(int percent)? onProgress,
  }) async {
    await initModel(model);

    final Whisper whisper = Whisper(model: model);
    final DateTime start = DateTime.now();
    const bool translate = false;

    try {
      final WhisperTranscribeResponse transcription = await whisper.transcribe(
        transcribeRequest: TranscribeRequest(
          audio: audioPath,
          language: lang,
          isTranslate: translate,
          isNoTimestamps: !withSegments,
          splitOnWord: splitOnWord,
          isRealtime: true,
          diarize: diarize,
          initialPrompt: initialPrompt,
          noContext: noContext,
          suppressNonSpeechTokens: suppressNonSpeechTokens,
          keepModelLoaded: keepModelLoaded,
        ),
        modelPath: _modelPath,
        onProgress: onProgress,
      );

      final Duration transcriptionDuration = DateTime.now().difference(start);

      return TranscribeResult(
        time: transcriptionDuration,
        transcription: transcription,
      );
    } catch (e) {
      debugPrint(e.toString());
      return null;
    }
  }

  /// Free the model parked in native memory by a transcription with
  /// `keepModelLoaded: true`. Safe to call when nothing is parked.
  Future<void> releaseModel() {
    // Which model the Whisper instance was created with is irrelevant here:
    // the native side frees whatever model is parked.
    return const Whisper(model: WhisperModel.tiny).releaseModel();
  }

  static Future<String> getModelDir() async {
    if (!Platform.isAndroid &&
        !Platform.isIOS &&
        !Platform.isMacOS &&
        !Platform.isWindows &&
        !Platform.isLinux) {
      throw UnsupportedError(
        'whisper_ggml supports Android, iOS, macOS, Windows, and Linux. '
        '${Platform.operatingSystem} has no native whisper implementation.',
      );
    }
    // getLibraryDirectory only exists on Apple platforms.
    final Directory libraryDirectory = Platform.isIOS || Platform.isMacOS
        ? await getLibraryDirectory()
        : await getApplicationSupportDirectory();
    return libraryDirectory.path;
  }

  /// Get local path of model file
  Future<String> getPath(WhisperModel model) async {
    _dir ??= await getModelDir();
    return '$_dir/ggml-${model.modelName}.bin';
  }

  /// Download [model] to [destinationPath]
  Future<String> downloadModel(WhisperModel model) async {
    if (!File(await getPath(model)).existsSync()) {
      final request = await HttpClient().getUrl(model.modelUri);

      final response = await request.close();

      final bytes = await consolidateHttpClientResponseBytes(response);

      final File file = File(await getPath(model));
      await file.writeAsBytes(bytes);

      return file.path;
    } else {
      return await getPath(model);
    }
  }
}