Voice AI that's
wide awake.
Frontier speech-to-text and text-to-speech — two real-time APIs for phone lines, clinical notes, and voice agents that respond like they're actually listening.
A transcript that arrives after the call is a record.
One that arrives mid-sentence is a conversation.
That difference — a few hundred milliseconds — is the entire product. It's what lets a nurse keep her eyes on the patient instead of a screen, and a caller forget they're not talking to a person. We build speech models that live inside that gap, and nothing else.
Two frontier models. Each its own API.
Streaming transcription that keeps up with real speech — 38ms median, speaker turns, and medical vocabulary out of the box.
Warm, natural speech that starts talking in 210ms — tuned for phone audio, with voices that hold up over an hour-long call.
Built for conversations with stakes.
Start with ten lines. Ship the same day.
Two websockets — one for Rem, one for Reverie. Plain JSON in, audio and text out. SDKs for JS and Python, or hit the sockets raw.