Everyone benchmarks time to first token. Almost everyone benchmarks it wrong for voice. That’s the uncomfortable truth in the latest deep-dive across the voice AI stack, and the reason is mechanical: a text-to-speech model can’t synthesize half a word. It needs a complete clause before it produces audio. So the metric that matters is what LiveKit calls time-to-first-sentence, or TTFS. You can have a blazing-fast first token and still have an agent that feels like it’s buffering mid-thought.

The latency math is unforgiving. LiveKit’s breakdown puts speech-to-text at 100-200ms, LLM streaming at 300-500ms, TTS at 100-200ms, and WebRTC network at 50-150ms. That lands most production systems between 700ms and 1.2 seconds end-to-end. Kwindla Hultman Kramer, the Pipecat co-creator, has gone further: target 800ms median voice-to-voice with 1,500ms acceptable for a rough proof of concept. Daily’s earlier research gives the human baseline — roughly 500ms is normal conversational response time, and anything past 800ms starts to feel off. That leaves about 700ms of TTFT budget for the LLM alone.

The benchmark numbers are where things get quietly damning. Mercury 2 generates an eye-popping 770 tokens per second, but its first chunk takes 3.07 seconds to arrive. That’s four times the entire LLM budget for natural conversation. It’s a throughput monster that would fail a voice agent outright. Cerebras and Groq tell a different story: respectable TTFT plus exceptional throughput, which is exactly the combination that makes time-to-first-sentence feel snappy. Even the same model behaves differently depending on who serves it — GPT-5.6 Luna non-reasoning runs 0.59 seconds on Bedrock but 0.74 seconds on OpenAI’s own API.

Methodology matters more than most teams admit. Artificial Analysis quietly changed its default workload in March 2026 from 1k to 10k input tokens, which raises both TTFT and output speed. That’s actually realistic for voice, since production agents front-load persona, policy, escalation rules, and tool schemas. Server location is baked into the numbers too — they test from Google Cloud’s us-central1-a. And Daily is blunt that TTFT varies substantially between runs, with providers swapping inference stacks without changing model names. If you’re picking a voice API on a single TTFT snapshot, you’re optimizing for a number that may not exist next week.