Release2026-10-27
AssemblyAI released AssemblyAI-TTS-1, a text-to-speech model for voice agents. It streams audio in 90 ms in production. The model uses continuous audio representations for natural speech. It costs $15 per 1 million characters. It offers unlimited concurrency with no annual contract. It is now generally available. The median first-byte latency is 47 ms. It was trained on millions of hours of conversational audio.
What changed
Text-to-speech models used discrete audio tokens, causing higher latency and lower quality.
What it unlocks
Streaming natural speech for live voice agents at under 100 ms latency.
- 90 ms streaming latency
- $15 per 1M characters
- 47 ms median first-byte latency
Sources