Grok's new audio APIs: Voice gets chopped into useful plumbing

xAI has released standalone Grok Speech to Text and Grok Text to Speech APIs, predictably doing the most useful thing possible: taking their shiny voice assistant and breaking it down into developer-sized pieces. According to xAI's announcement, these APIs run on the same stack powering Grok Voice, Tesla vehicles, and Starlink support. The headline isn't that a voice assistant can talk—we've been there. The headline is xAI turning voice into an infrastructure surface, complete with batch transcription, streaming, and expressive speech tags.

The feature list for speech-to-text focuses heavily on the practical: word-level timestamps, speaker diarization, multichannel support, and handling over 25 languages. It's unglamorous, which is exactly why it matters. Audio AI products usually die in the messy middle—figuring out who spoke when, or if a laptop microphone in a crowded room completely breaks the transcript. xAI priced batch transcription at $0.10 per hour and streaming at $0.20 per hour. If those numbers hold up in production, Grok STT enters true infrastructure territory. Audio storage anxiety is real, and transcription costs add up fast.

On the text-to-speech side, xAI offers REST generation for long-form audio and WebSocket streaming for real-time applications. But the genuinely interesting part is the support for expressive speech tags. Developers can inject a `[laugh]`, `[sigh]`, `[whisper]`, or adjust emphasis and pacing. This makes voice systems productively dangerous. A steerable voice is vastly better for education or accessibility than a flat robotic narrator, though we fully expect someone to build a whispering customer-service bot. A modular API lets developers use STT for transcription, TTS just for alerts, or combine them without buying into a locked-in voice agent dream. xAI is selling the plumbing, and for anyone building real products, that's a massive win.