voice_forge_flutter #

Flutter client for voice_forge voice agents: mic → WebRTC → agent, with the agent.events data-channel contract, automatic playback of agent audio, and instant barge-in.

Pure Dart on top of flutter_webrtc + web_socket_channel — works on web, iOS, Android, and desktop. No LiveKit, no tokens.

Features #

• One-tap voice calls: VoiceCallController drives mic capture, WebRTC, signaling, and playback.
• Data-channel contract (agent.events): user transcripts, assistant text, agent state, call summaries, and booking confirmations arrive as typed events.
• Barge-in: sendBargeIn() interrupts the agent the instant you start talking; the controller also exposes the raw mic stream (localStream) so you can add your own local speech-onset detection.
• RTT probes: built-in ping/pong on the data channel surfaces round-trip times via rttMs.
• Clean teardown (end() / dispose()).

Getting started #

This package is the client half: it connects to a running voice_forge agent server (any process speaking the voice_forge signaling contract over WebSocket at ws://host:8765/signal — typically a voice_forge server with an AgentSession loop). See the voice_forge repository for a complete agent server example: https://github.com/iamudesharma/clinicguard

Add the dependency:

dependencies:
  voice_forge_flutter: ^0.1.0

Usage #

import 'package:voice_forge_flutter/voice_forge_flutter.dart';

final call = VoiceCallController(
  signalingUrl: 'ws://localhost:8765/signal',
  // iceServers: optional, defaults to Google's public STUN
);

call.phase.listen((p) => print('phase: $p')); // idle/connecting/connected/error
call.events.listen((e) {
  switch (e['type']) {
    case 'user_transcript':
      print('you: ${e['text']}');
    case 'assistant_text':
      print('agent: ${e['text']}');
    case 'agent_state':
      print('state: ${e['state']}'); // idle/listening/thinking/speaking
    case 'summary':
      print('summary: ${e['text']}');
  }
});
call.rttMs.listen((ms) => print('data-channel RTT: ${ms}ms'));

await call.start();     // mic -> agent; agent audio plays automatically
call.sendBargeIn();     // interrupt the agent instantly
await call.send({'event': 'patient_id', 'patient_id': 'PAT-123'});
await call.setMicEnabled(false); // mute without hanging up
await call.end();

The events stream carries the full voice_forge contract — anything the agent publishes on agent.events (user_transcript, assistant_text, agent_state, summary, booking_confirmed, connected, …) is forwarded as-is, and any non-JSON message arrives as {'type': 'raw', 'text': ...}.

Additional information #

• Server framework: voice_forge (pure Dart)
• Repository: https://github.com/iamudesharma/clinicguard (set to the real URL before the first release)
• License: MIT