Text-to-speech, streamed.
A fine-tuned Shona voice, delivered either as a finished WAV or as raw PCM over a WebSocket with barge-in — so an agent can stop mid-sentence when the caller speaks.
A voice tuned for one accent, not forty
TeeVoice Speak runs a fine-tuned checkpoint rather than a general multi-voice model — Natasha, a Shona female voice, is the production default.
Everything Text-to-Speech API does
Audio frames arrive as they are produced, so playback starts before the sentence is finished.
One clear frame drops buffered and in-flight audio, so the agent stops the moment the caller speaks.
The synchronous route returns a finished WAV on a CDN URL; the socket returns raw frames and stores nothing.
Quota counts seconds of audio actually produced — cut a call short and you pay for what was spoken.
{"type":"text","text":"Your balance is …"}{"type":"clear"}{"type":"text","text":"Go ahead — I'm listening."}clear, it talks over them until the buffer drains.Built to be interrupted
Drive the socket with text, flush, clear and end frames. Unknown frame types are ignored rather than fatal, so a forward-compatible client cannot kill its own stream.
Listen. Reason. Speak. In real time. One AI platform for every conversation.