Grok Voice is the voice part of the Grok API from SpaceX Super Intelligence LLC (SpaceXSI), the company at x.ai that makes the Grok models. It has three services, all powered by Grok: speech-to-speech voice agents, text-to-speech and speech-to-text.
The speech-to-speech API runs full-duplex conversations, in which both sides can speak at once and the caller can interrupt, over a WebSocket connection to the /v1/realtime endpoint with sub-second latency. During a call the agent can reason, call tools and search the web; client-side apps connect with short-lived ephemeral tokens instead of the API key. Text-to-speech offers more than 25 voices in over 25 languages, speech tags for tone, pauses, whispers and laughter, and formats such as PCM, MP3 and WAV. Speech-to-text transcribes files and live streams, with speaker diarization (labeling who said what). Custom voices can be cloned from two minutes of speech.
It is a hosted API billed by usage: per minute of conversation, per million characters of speech and per hour of transcribed audio. Enterprise options include zero data retention.