AssemblyAI has released AssemblyAI-TTS-1, its first text-to-speech model, now generally available for voice agents at $15 per 1 million characters. Why it matters: a company long known for speech-to-text now sells speech out as well, so developers can run a full phone conversation through one vendor instead of stitching two together.
October 29, 2026
What happened
AssemblyAI said TTS-1 is built specifically for live phone calls, where latency and pronunciation decide whether a caller stays on the line. The model is hosted only, in the US (us-east-1) or EU (eu-north-1), with no open weights or on-premise option. It is available through an HTTP API, a WebSocket API, and an OpenAI-compatible route, with reference integrations for LiveKit and Pipecat.
Pricing is a flat $15 per 1 million characters, including spaces and punctuation. AssemblyAI says every voice, cloning, word timestamps, and EU processing are included at that rate, with no annual contract, no minimums, and unlimited concurrency on pay-as-you-go.
The details that matter for builders
The launch library includes more than 4,000 voices, and every voice speaks all six supported languages: English, Spanish, French, German, Italian, and Portuguese. Instant voice cloning needs a short recording and confirmation of the speaker’s permission; the clone in the launch post used 24 seconds of audio.
Unlike most text-to-speech models, which represent audio as discrete tokens, AssemblyAI says TTS-1 works on continuous audio representations, trained on millions of hours of audio including real conversational data. Output is native 24 kHz PCM, with mu-law and A-law encodings for telephony.
On AssemblyAI’s own tests, dated October 28, TTS-1 scored 93.7% on entity accuracy — reading back phone numbers, codes, and similar hard targets — just ahead of ElevenLabs v4 at 93.6%, with a 47 ms median time to first byte. Independent results from Cekura were more mixed: 159 ms median streaming latency, 81.5% reading accuracy, and a 4.50 naturalness score across its benchmark.
At launch the model does not yet support SSML, speed control, pronunciation dictionaries, or emotion tags.
Why it matters
Voice agents live or die on the seams between vendors. If your transcription already runs on AssemblyAI, adding its text-to-speech removes one integration, one bill, and one latency hop. The headline price also undercuts better-known rivals on list price, though at least one competitor matches $15 on demand — the sharper difference AssemblyAI is selling is unlimited concurrency without buying a tier. For teams testing a phone agent today, that makes TTS-1 a straightforward trial: no contract, and the benchmarks, both the company’s and the independent one, are published for you to check against your own calls. That is the Tested Intelligence take: judge it on your own order numbers and account codes, not on a demo script.
Frequently asked questions
How much does AssemblyAI-TTS-1 cost?
$15 per 1 million characters, with no annual contract, no per-voice fees, and unlimited concurrency included.
What languages does AssemblyAI-TTS-1 support?
Six at launch: English, Spanish, French, German, Italian, and Portuguese. Every voice in the library speaks all six.
Can AssemblyAI-TTS-1 clone a voice?
Yes, instant cloning is included. It needs a short recording and confirmation that the speaker has given permission.
Sources: MarkTechPost, AssemblyAI