Time to First Audio (TTFA)

Time to First Audio is the wall-clock latency from end-of-user-utterance to first-audio-out. It is the sum of: end-of-turn detection delay (typically 300 to 500ms after the user stops speaking), STT to final transcript (~150ms), LLM time-to-first-token (~200 to 400ms with streaming), and TTS time-to-first-chunk (~150 to 250ms). Total: 800ms to 1300ms in production. Below 800ms feels conversational. Above 1500ms feels broken. TTFA is the metric users perceive as "the agent's response time," and it is the single biggest determinant of whether a voice agent feels alive or dead.

Two optimizations dominate the TTFA fight. First, streaming everywhere: STT streaming partials to the LLM, LLM streaming tokens to the TTS, TTS streaming audio chunks to the caller. Without streaming, your TTFA is the SUM of full-stage latencies, which is impossible to keep under 2 seconds. Second, model selection: a Haiku-class model at 200ms time-to-first-token is dramatically better for voice than an Opus-class model at 800ms, even if Opus is smarter. The right pattern is small fast LLM for voice, escalate to larger model only when needed.

Track TTFA on every voice call. Set an alert when the p95 exceeds your target (typically 1000ms for B2C, 1500ms for B2B). Optimize the slowest stage first.

• LiveKit Agents performance
• Pipecat metrics docs

Related Terms