AssemblyAI-TTS-1 is the company's first text-to-speech model, built for developers whose voice agents have to answer on a live call. AssemblyAI quotes about 90 milliseconds to first audio in production, and in its own lab comparison the median time to first byte was 47 milliseconds, the fastest of the six models it tested. The list price is $15 per million input characters, with unlimited concurrency on pay-as-you-go and no annual contract, and that rate includes the voice catalog, instant cloning, word timestamps, and processing in the US or the EU. AssemblyAI trained the model on millions of hours of audio, including real conversations, for calls where a slow or flat reply gets the caller to hang up. The catalog has more than 4,000 voices, and developers can clone a voice from a short clip. English, Spanish, French, German, Italian, and Portuguese ship at launch, with more languages planned. A WebSocket session can stay open for the whole conversation so text can be sent as a language model produces it, while HTTP and OpenAI-compatible endpoints cover one-shot requests, and the docs include Pipecat and LiveKit integrations. The launch video puts the model on a pizza-order call, including a child trying to place the order, where quantities and extras have to come back clearly. On AssemblyAI's pronunciation set, the model scored 93.7% entity accuracy on hard targets such as phone numbers, dates, and confirmation codes. Listeners on Artificial Analysis ranked it 10th in the controlled-voice arena and 12th in the provider-voice arena. AssemblyAI was founded in 2017 by Dylan Fox and already sells speech-to-text through a developer API, so this release adds a speaking model on the same platform teams use for transcription.