ai_live_voice

Flutter package for real-time AI voice conversations using the Gemini Live API: microphone capture, speaker playback, voice barge-in, and optional tool calling.

Supported platforms (v1)

iOS, Android, macOS, Windows, Linux. Web is not supported in v1.

Quick start

dependencies:
  ai_live_voice: ^0.0.1-dev.1

import 'package:ai_live_voice/ai_live_voice.dart';
import 'package:ai_live_voice/gemini_tools.dart';

final client = AILiveVoiceClient(
  apiKey: const String.fromEnvironment('GEMINI_API_KEY'),
  onToolCalls: myHandler, // required if `liveTools` is non-empty
);

client.status.listen((status) {
  // VoiceIdle, VoiceConnecting, VoiceListening, VoiceSpeaking, VoiceError
});

await client.start();
// ...
await client.stop();
await client.dispose();

Run with:

flutter run --dart-define=GEMINI_API_KEY=your_key

Optional model override:

flutter run --dart-define=GEMINI_API_KEY=... --dart-define=GEMINI_LIVE_MODEL=gemini-2.0-flash-live-preview-04-09

Example app

A minimal evaluator app lives in example/:

cd example
flutter run --dart-define=GEMINI_API_KEY=your_key

Tool calling

Define tools with types from gemini_tools.dart (Gemini-specific in v1). Implement VoiceToolCallHandler and pass it to AILiveVoiceClient when liveTools is non-empty.

Host responsibilities

• Provide a valid API key at runtime (never commit keys).
• Add microphone permission strings / entitlements on each platform.
• Own UI for start/stop and error display (AILiveVoiceClient.status only).
• Reuse one AILiveVoiceClient per tool set — config is fixed for the instance lifetime.

Voice barge-in

While the model speaks, real mic audio is sent so server VAD can interrupt. On interrupted, queued speaker audio is cleared. A short post-turn silence uplink (~450 ms) reduces speaker→mic false triggers (v1 uses this profile on all outputs).

License

MIT — see LICENSE.