What is AssemblyAI-TTS-1?
AssemblyAI-TTS-1 is the transcription company's first text-to-speech model, released on 27 October 2026 and generally available from launch. It is built for voice agents on live calls: it streams audio, returns word timestamps and accepts a cancel message when the caller interrupts.
The library has more than 4,000 voices, and instant cloning makes a new voice from a short recording in a few seconds once you confirm the speaker's consent. It speaks English, Spanish, French, German, Italian and Portuguese, with more promised. AssemblyAI says the model was trained on millions of hours of audio, including real conversations, and works on continuous audio rather than discrete tokens.
One price covers every voice, cloned voices, timestamps and processing in either the US or the EU, with no concurrency tiers or minimums.
How much does AssemblyAI-TTS-1 cost?
| Plan | Price | What you get |
|---|---|---|
| Pay as you go | $15 per 1M characters | All voices, cloning, US or EU processing, unlimited concurrency |
| Committed volume | Custom | Discounts for committed usage |
$15 per 1M input characters (spaces and punctuation count), including cloned voices, word timestamps and EU processing; no minimums. Volume discounts on commitment.
Is AssemblyAI-TTS-1 worth it?
AssemblyAI-TTS-1 is priced to be the easy second purchase for teams already using AssemblyAI to listen. At $15 per million characters, about 81 cents per hour of finished speech, it costs well under ElevenLabs' $40 and Deepgram's $30 floor prices, as AssemblyAI frames it, and the rate does not change for cloned voices, EU processing or ten thousand concurrent calls.
Its strengths are the dull ones that matter on a call. In Cekura's independent benchmark it ranked first on reading phone numbers and codes, at 91.7%, and fourth on naturalness. AssemblyAI's own test puts its accuracy on names, numbers and dates at 93.7%, level with ElevenLabs v4 at a fifth of the price. Artificial Analysis's listener arena places it 20th, with 1,212 Elo, a respectable debut rather than a leading one.
The catch: the latency figures disagree, and the feature list is short. AssemblyAI quotes 90 ms to first audio in production and 47 ms in a single-client lab run; Cekura measured 159 ms, eighth of 19. There is no SSML, no speed setting, no pronunciation dictionary and no emotion tags, only six languages, and no fine-tuned professional cloning yet.
Our take: this is a sensible voice for an agent that reads back order numbers to callers in Western Europe. For an audiobook or a brand voice that has to act, the same $15 buys more expressive models elsewhere on the leaderboard.
Where it earns its place
A car-insurance claims line handling 2,000 calls a day
Read back policy numbers, dates and addresses without the caller asking twice
Pair TTS-1 with STT-3.7-Realtime-Agent through the LiveKit reference agent. At about 81 cents per hour of speech, 1,000 hours of agent speech a month cost about $810.
A European bank's contact-centre team
Keep synthetic speech inside the EU for compliance
Point the streaming client at the EU host; the voice library and cloning work there at the same $15 per million characters.
Before you commit
- Setup
- Stream text over the TTS WebSocket or call the HTTP endpoint with a library voice ID; clone a voice by uploading a recording and confirming consent, ready in seconds.
- Integrations
- AssemblyAI API with US and EU hosts; reference agents for LiveKit and Pipecat; word timestamps and a cancel message for interruptions.
- Data and privacy
- No published weights. AssemblyAI's data policy covers audio sent to its APIs; check it before cloning staff or customer voices.
- Cost at scale
- A million characters is roughly 18 hours of audio at 15 characters a second, so 10,000 hours of speech cost about $8,100 before any volume discount.
- Limits
- English, Spanish, French, German, Italian and Portuguese only. No SSML, speed control, pronunciation dictionaries, emotion tags or professional cloning at launch.
What’s good
- $15 per 1M characters flat, including clones and EU processing
- First on phone numbers and codes in Cekura's benchmark
- Unlimited concurrency on pay-as-you-go, no contract
- Word timestamps and barge-in cancel for agents
What’s not
- No SSML, speed control, pronunciation dictionary or emotion tags
- Six languages at launch
- 20th on Artificial Analysis's arena, Elo 1,212
- Independent latency of 159 ms, far from the 47 ms lab figure
Best for
Teams running phone voice agents in English or major Western European languages, especially those already on AssemblyAI's speech-to-text.
AssemblyAI-TTS-1 alternatives
AssemblyAI-TTS-1: common questions
- Is AssemblyAI-TTS-1 worth it?
- Fast, fairly priced and good at reading phone numbers aloud, but a first release that still lacks the controls a production voice team expects. We rate it 3.9 out of 5.
- How much does AssemblyAI-TTS-1 cost?
- $15 per 1M input characters (spaces and punctuation count), including cloned voices, word timestamps and EU processing; no minimums. Volume discounts on commitment.
- What are the downsides of AssemblyAI-TTS-1?
- No SSML, speed control, pronunciation dictionary or emotion tags; six languages at launch; 20th on Artificial Analysis's arena, Elo 1,212; independent latency of 159 ms, far from the 47 ms lab figure.
- Who is AssemblyAI-TTS-1 best for?
- Teams running phone voice agents in English or major Western European languages, especially those already on AssemblyAI's speech-to-text.
- How much does AssemblyAI-TTS-1 cost for a team or in production?
- A million characters is roughly 18 hours of audio at 15 characters a second, so 10,000 hours of speech cost about $8,100 before any volume discount.
- Is AssemblyAI-TTS-1 safe for company data?
- No published weights. AssemblyAI's data policy covers audio sent to its APIs; check it before cloning staff or customer voices.
- What are the best AssemblyAI-TTS-1 alternatives?
- The closest alternatives we have reviewed are Inworld Realtime TTS-2 and TTS-2 Flash, Cartesia Sonic 3.6, Eleven v4 and v4 Turbo and Speechify Simba 3.5.
Answers checked Nov 21, 2026.