AssemblyAI announced that AssemblyAI-TTS-1 is generally available, calling it its first text-to-speech model and positioning it for voice agents on live calls. The company quotes 90 ms to first audio in production, a figure it describes as the conservative end of what it sees across real networks and two regions.
AssemblyAI lists the price at $15 per million input characters. It says its library has more than 4,000 voices, the model ships with instant voice cloning, and developers need no spend commitment or concurrency tier. At launch, the company said the model supported various accents in English, Spanish, French, German, Italian and Portuguese. “Pro” voice cloning, custom pronunciation dictionaries and emotion tagging were not yet supported.
The production latency claim differs from the 47 ms median time to first byte that AssemblyAI measured in a separate single-client comparison. Artificial Analysis reported an 81.7% overall score and a No. 16 rank for the model on its Pronunciation Robustness benchmark, which uses human reviewers to judge challenging text across four categories. It also reported generation at 162 characters per second, faster than five models it named, including Cartesia Sonic 3.6 at 148.6 characters per second. That generation-speed comparison does not measure time to first audio. Artificial Analysis listed Speechify Simba 3.2 at $6.6 per million characters, below AssemblyAI’s $15 list price.
Comments