Streaming TTS

Stream TTS audio as it’s generated for faster time-to-first-audio, especially with longer text.

Overview

Streaming TTS starts playing audio before the entire synthesis is complete. This is particularly useful for:

Long text passages
Voice assistants responding in real-time
Reducing perceived latency

synthesizeStream

synthesizeStream yields TTSOutput chunks as the backend produces them. Each chunk carries its own audioData, chunkIndex, and isFinal flag.

await for (final chunk in RunAnywhere.synthesizeStream(longText)) {
  // chunk.audioData is raw PCM in chunk.audioFormat
  appendToPlaybackBuffer(chunk.audioData);
  if (chunk.isFinal) break;
}

The signature is synthesizeStream(String text, {TTSOptions? options}), with options named rather than positional.

Failure contract

The stream never throws. An uninitialized SDK, a failed phase-2 readiness check, or no loaded voice finishes the stream silently with no chunks. A session-start failure arrives as one terminal TTSOutput with isFinal true and a non-empty errorMessage prefixed TTS stream failed:.

if (chunk.isFinal && chunk.errorMessage.isNotEmpty) {
  showError(chunk.errorMessage);
}

Chunked Synthesis

If you want sentence-level boundaries rather than backend chunks, split the text yourself and call speak per sentence. speak synthesizes and plays through the SDK’s playback manager, so each call blocks until its audio is queued.

Future<void> speakInChunks(String longText) async {
  final sentences = longText.split(RegExp(r'(?<=[.!?])\s+'));
  for (final sentence in sentences) {
    await RunAnywhere.speak(sentence);
  }
}

With Voice Agent Pipeline

The voice agent runs synthesis and playback for each turn itself. Load the three models, initialize the agent, and consume events.

await RunAnywhere.stt.load('sherpa-onnx-whisper-tiny.en');
await RunAnywhere.llm.load('smollm2-360m-q8_0');
await RunAnywhere.tts.loadVoice('vits-piper-en_US-lessac-medium');

await RunAnywhere.initializeVoiceAgentWithLoadedModels();

RunAnywhere.streamVoiceAgent().listen((event) {
  switch (event.whichPayload()) {
    case VoiceEvent_Payload.userSaid:
      print('User: ${event.userSaid.text}');
    case VoiceEvent_Payload.assistantToken:
      stdout.write(event.assistantToken.text);
    case VoiceEvent_Payload.audio:
      print('Playing response...');
    case VoiceEvent_Payload.agentResponseCompleted:
      print('Ready for next turn');
    default:
      break;
  }
});

Latency Optimization Tips

Preload Voice Model

Load the TTS voice during app startup or idle time, not when the user first needs it.

// In app initialization
await RunAnywhere.tts.loadVoice('vits-piper-en_US-lessac-medium');

Use Faster Voices

Smaller voice models synthesize faster. Choose based on your quality/speed tradeoff.

Sentence-Level Streaming

For very long responses, synthesize and play sentence by sentence rather than waiting for the complete response.

Buffer Management

Start playback as soon as you have enough audio buffered (typically 100-200ms).

synthesize()

Basic synthesis

Voice Agent

Complete voice pipeline

Getting Started

Swift SDK

Kotlin SDK

React Native SDK

Flutter SDK

Web SDK

Vibe Coding

Overview

synthesizeStream

Failure contract

Chunked Synthesis

With Voice Agent Pipeline

Latency Optimization Tips

See Also

synthesize()

Voice Agent

​Overview

​synthesizeStream

​Failure contract

​Chunked Synthesis

​With Voice Agent Pipeline

​Latency Optimization Tips

​See Also

synthesize()

Voice Agent

Overview

synthesizeStream

Failure contract

Chunked Synthesis

With Voice Agent Pipeline

Latency Optimization Tips

See Also