The last mile of a voice agent is turning its answer back into speech. For years that step was either slow or robotic. In 2026, the best text-to-speech models are both fast and genuinely expressive.
Time-to-first-byte is the metric
For conversational TTS, the number that matters most is time-to-first-byte — how quickly the first chunk of audio starts playing. The leading real-time models now hit well under 200ms, with the fastest around 150ms, and stream the rest as they go. That means the agent can start speaking almost immediately rather than generating the whole clip before playback.
A voice agent that pauses to render its full reply feels broken. One that starts speaking in 150ms feels alive.
Beyond speed: prosody
Fast isn't enough — the audio has to sound human. The advance is in prosody: natural rhythm, emphasis, and intonation that match the meaning, plus conversational touches like appropriate pauses. Modern models generate this expressiveness while streaming, so responsiveness and naturalness no longer trade off against each other.
What to weigh
When choosing a TTS model: latency (time-to-first-byte and streaming), voice quality and expressiveness, voice options and cloning, language coverage, and cost. As always, benchmarks are a starting point — the real test is whether a specific voice sounds right for your use case, which only listening will tell you. Confirmed: sub-200ms streaming synthesis is now routine. The rest is taste and fit.
