Grok Voice Transcribe 2.0

Grok Voice Transcribe 2.0 is a speech-to-text model from xAI, whose documentation and website now give the company name as SpaceXAI. It was announced on 18 September 2026 and is available through the Speech to Text API as grok-voice-transcribe-2.0. The company says it is built on the audio foundation model behind Grok Voice and trained on live, noisy, multilingual audio.

The API accepts recorded files or URLs (up to 500 MB, in 12 audio formats) and real-time streams over WebSocket, including Opus packets for low-bandwidth clients. Features include word-level timestamps, speaker diarisation at no extra charge, up to eight independently transcribed channels, biasing towards up to 100 key terms, filler-word removal, smart turn detection for voice agents, and written-form formatting of numbers and currencies in 25 languages. The model detects the spoken language automatically.

Pricing is unchanged from version 1.0: $0.10 per hour of audio for batch transcription and $0.20 per hour for streaming, with diarisation, timestamps and key terms included.

In its own evaluation, xAI reports that version 2.0 is about twice as accurate as 1.0 and that word error rate on a 19-language short-phrase set fell from 20.6% to 6.8%. It also says the model ranked first for accuracy among 32 streaming models on the Artificial Analysis leaderboard at launch. MarkTechPost, reporting the release on 18 September, noted that these results are vendor-reported. Atlassian is named as a launch customer for Loom transcription.

The documentation is inconsistent about the default: the speech-to-text page says 2.0 is used when no model is given, while the release notes say 1.0 remains the default. xAI says 1.0 will be deprecated in the coming weeks, so integrations should name a model explicitly.