Overview
AssemblyAI-TTS-1 is AssemblyAI's first text-to-speech model, generally available since October 27, 2026. The company built its business on speech-to-text and now covers the speaking half, selling a complete conversational loop: hearing, deciding and replying.
It optimises for one setting, voice agents on phone calls, and three numbers carry the pitch: 90 milliseconds to first audio, 15 dollars per million characters, and unlimited concurrency on pay-as-you-go with no annual contract. AssemblyAI's own lab measurement puts median time to first byte at 47 milliseconds, the fastest of six models tested from a single client.
The architecture makes one notable trade. Many speech models quantise audio into discrete tokens before generating; AssemblyAI operates directly on continuous audio representations, which it says improves both realism and time to first sound. Generation runs at 162 characters per second, ahead of the 148.6 characters per second of the fastest model in its comparison set.
Key Features
- 90 milliseconds to first audio: Production time to first audio is 90 milliseconds, with a 47 millisecond median first byte in the lab. That is the difference between a natural pause and the moment a caller starts pressing zero for an operator.
- Unlimited concurrency, no annual contract: Pay-as-you-go has no concurrency ceiling. Ten simultaneous calls and ten thousand run through the same API at the same rate, with no tier to buy and no annual commitment.
- More than 4,000 voices: The library exceeds 4,000 voices with instant cloning; the published demo clones from 24 seconds of audio. Coverage spans English, Spanish, French, German, Italian and Portuguese, each with multiple accents.
- Streams from the first byte: Audio streams as it generates, a cancel message handles barge-in when the caller interrupts, and word timestamps come back with the audio, so an existing turn loop needs no extra buffering layer.
- Continuous audio representations: Instead of discretising audio into tokens, the model works on continuous representations. AssemblyAI credits this with both the naturalness and the 47 millisecond first byte.
- Two regions and reference agents: US and EU data residency at the same price. Reference agents for LiveKit and Pipecat let teams swap the TTS stage of an existing pipeline.
Use Cases
- Inbound and outbound phone agents where 90 milliseconds keeps callers on the line.
- High-volume short calls such as booking, reminders and follow-ups, where per-character billing tracks usage better than a subscription.
- Compliance work needing US or EU data residency at a single rate.
- Teams already on LiveKit or Pipecat, replacing the TTS stage only.
- Multilingual announcements across six languages with accent variants at one rate.
- Brand-specific voices through instant cloning from a 24 second sample.
Pros
- 90 milliseconds to first audio in production, 47 millisecond median first byte in the lab.
- 15 dollars per million characters, the same for every voice and language.
- Unlimited concurrency on pay-as-you-go with no annual commitment.
- Over 4,000 voices with cloning from a 24 second sample.
- Streams from the first byte with barge-in cancellation and word timestamps.
- US and EU regions at one price, with LiveKit and Pipecat reference agents.
Pricing
15 dollars per million characters, billed on characters sent including spaces and punctuation. All voices, languages, instant cloning, word timestamps and both data regions are included at that rate. Concurrency is unlimited on pay-as-you-go, with committed-volume pricing available separately. Finer pronunciation control and SSML-style features are on the published roadmap; check the documentation for current availability.
Summary
The speech synthesis market is crowded with more natural voices, so TTS-1 competes on price and commercial terms instead. For teams building phone agents the question is whether each call makes money, not whether the mean opinion score is a tenth of a point higher. Six languages is a narrow footprint, though, so check that your target languages are in the first wave before committing.
Version History
- AssemblyAI-TTS-1 generally available: 90 milliseconds at 15 dollars per million characters (2026-10-27): AssemblyAI released its first text-to-speech model, AssemblyAI-TTS-1, with 90 milliseconds to first audio in production and a 47 millisecond median first byte in lab tests, generating at 162 characters per second. Pricing is 15 dollars per million characters with unlimited concurrency on pay-as-you-go and no annual contract, with US and EU data residency at the same rate. The library exceeds 4,000 voices with instant cloning from 24 seconds of audio, covering English, Spanish, French, German, Italian and Portuguese. Streaming, barge-in cancellation and word timestamps are supported, with LiveKit and Pipecat reference agents. Artificial Analysis ranks it 12th in the Provider Voice Arena at 1227 Elo, 10th in Controlled Voice, and 16th on pronunciation robustness at 81.7 percent.