Thanks — I’m responding to your request about implementing fallbacks for STT and TTS failures. Short answer: you can use built‑in fallback for STT and implement a small custom node for TTS fallback. Below are concise steps and examples. 1) STT fallback (built‑in) • Configure a primary model with a

fallback

array. The SDK will try the next model if the primary fails.

import { inference } from '@livekit/agent-sdk';

const stt = new inference.STT({
  model: 'deepgram/nova-3',
  fallback: ['assemblyai/universal-streaming'],
});

2) TTS fallback (custom) • There’s no automatic

fallback

array for TTS. Override the TTS node to catch errors and call a secondary provider or return a pre-recorded clip.

import { TTSNode } from '@livekit/agent-sdk';

class MyTTSNode extends TTSNode {
  async synthesize(text: string) {
    try {
      return await super.synthesize(text); // primary
    } catch (e) {
      console.warn('Primary TTS failed, falling back:', e);
      return await fallbackProvider.synthesize(text); // secondary
      // or: return await this.loadAudio('fallback.mp3');
    }
  }
}

3) UX and error handling recommendations • For STT low-confidence/empty results: prompt the user to repeat before passing input downstream. • For TTS failures: switch to a simpler provider or play a short pre-recorded message to avoid dead air. 4) Observability & testing • Use Agent Observability to track per-stage errors and configure retries. • Test by forcing failures (invalid key/network block) to verify fallback behavior. If you want, tell me: • which language/SDK you’re using (Node/Python/etc.), • which STT/TTS providers you plan to use, and I’ll give a tailored code sample and testing checklist. Sources: Sequential Pipeline Architecture for Voice Agents | LiveKit | Building Performant Voice Agents in India | LiveKit | Build Your First AI Voice Agent in Python: Complete Tutorial | LiveKit | Pipeline nodes and hooks | LiveKit Documentation, with interactive elements