voice_forge 0.3.0 copy "voice_forge: ^0.3.0" to clipboard
voice_forge: ^0.3.0 copied to clipboard

Published 28 days ago
SDKDartFlutter
PlatformAndroidiOSLinuxmacOSWindows
→

Metadata

Minimal voice-agent framework for Dart: WebRTC transport, speech (VAD/STT/TTS), LLM interface and the conversation loop — all pure Dart.

More...

0.3.0 #

  • Streaming LLM replies: VoicepipeLlm.streamReplyWithTools streams SSE content deltas to onPartial as they arrive and assembles tool_calls deltas by index; providers that ignore stream fall back to plain JSON. FallbackLlm and EchoLlm support it; per-call maxTokens (default 80) bounds responses.
  • Sentence-incremental TTS: AgentSession splits replies on sentence ends and synthesizes/speaks each sentence as it completes — first audio starts after the first sentence instead of the full reply, with synthesis of the next sentence overlapping playback of the current one.
  • Worker-isolate STT/TTS: SherpaKit.createWorkerSpeech() runs transcription and synthesis in a long-lived isolate (its own sherpa recognizer + TTS), so the event loop — and the instant-onset barge-in gate — stays live during speech work. Falls back to the main-isolate implementations on init failure.
  • Breaking (async speech interfaces): VoicepipeSTT.transcribe and VoicepipeTTS.synthesize now return Future (Future<String> / Future<TtsAudio>).
  • Adaptive turn detection: merge window shrinks to 200 ms for long, self-contained segments (was a fixed 450 ms); minSilenceDuration defaults to 0.30 s.
  • Latency cuts: LLM history capped at the last 8 user/assistant turns (system context always kept); greet(preSynthesized:) plays pre-rendered audio instantly; barge-in onset gate arms only while audio is actually playing (a reply is no longer cut by the tail of the user's own utterance).
  • Measured per-turn latency improvement of ~43% on the ClinicGuard agent (speech end -> first audio: ~3.3 s vs ~5.7 s with a free-tier LLM).

0.2.1 #

  • pub.flutter-io.cn score fixes: added an example/, shortened the pubspec description, and resolved all analyzer infos (async VoicepipeSTT, null-aware elements, formatting).

0.2.0 #

  • Renamed from voicepipe to voice_forge (same package, new brand; part of the *_forge family alongside image_forge_core / video_forge).
  • First-run auto-download: SherpaKit.load fetches the prebuilt libsherpa-onnx-c-api native library (into ~/.cache/voice_forge/native/, overridable via VOICE_FORGE_NATIVE_DIR) and any missing standard speech models (silero VAD, Whisper, Piper) from the official sherpa-onnx releases — no manual download step. Opt out with autoDownload: false.
  • SherpaKit.load / loadNative are now async (await SherpaKit.load(...)).
  • Initial pub.flutter-io.cn-ready release of the voice_forge framework.
  • WebRTC transport (VoiceCallServer, PeerSession) with WebSocket signaling and the agent.events data channel.
  • Speech stack via voice_forge_speech: Silero VAD, Whisper STT, Piper TTS.
  • LLM layer: OpenAiCompatibleLlm, FallbackLlm (provider chaining with cooldowns), offline EchoLlm, and OpenAI-style tool calling.
  • AgentSession conversation loop with barge-in, segment merging, event streams, and streaming TTS audio.

← Metadata

0
likes
150
points
169
downloads

Documentation

API reference

Publisher

unverified uploader

Weekly Downloads

Metadata

Minimal voice-agent framework for Dart: WebRTC transport, speech (VAD/STT/TTS), LLM interface and the conversation loop — all pure Dart.

Repository (GitHub)
View/report issues

Topics

#voice #webrtc #speech #llm #audio

License

MIT (license)

Dependencies

archive, http, opus_codec_dart, shelf, shelf_web_socket, voice_forge_speech, web_socket_channel, webrtc_dart

More

Packages that depend on voice_forge

Back