speech_core #

Provider-neutral speech recognition, synthesis, VAD, end-of-utterance, and diarization contracts.

Providers expose stable string IDs plus typed capability, model, voice, request, result, and failure objects. Provider SDK objects and untyped payload maps stay inside adapter packages.

import 'package:speech_core/speech_core.dart';

final registry = SpeechProviderRegistry();
registry.register(provider);

Streaming STT, VAD, and EOU sessions are audio sinks. TTS is an AudioSource, so synthesized PCM can branch to playback, recording, metering, or analysis.

TurnCompletionScorer scores a finite window for whether the speaker actually finished, which silence alone cannot tell you, and TranscriptInverseTextNormalizer rewrites a transcript to written form without discarding its word timings.

Available adapters in this repository include FluidAudio, MLX Audio, Deepgram, and OpenAI TTS.