Grok Voice Transcribe 2.0

xAI's speech-to-text model, Grok Voice Transcribe 2.0, was released on 18 September 2026 and can be called through the Speech to Text API using the model name grok-voice-transcribe-2.0. The company's documentation now gives its name as SpaceXAI.

Main features

• Batch and live transcription: recorded files, or links to them, up to 500 MB in 12 audio formats, plus live streams over WebSocket, with Opus packets supported for low-bandwidth clients.
• Speaker and timing detail: word-level timestamps, speaker diarisation and up to eight channels, each transcribed independently.
• Vocabulary control: biasing towards up to 100 key terms, with filler-word removal and smart turn detection for voice agents.
• Language handling: automatic language detection, and written-form formatting of numbers in 25 languages.

Customers and markets

Developers building voice agents and transcription features. Atlassian is named as a launch customer, using it for Loom transcription.

Key facts

• Accuracy claims: the vendor says version 2.0 is about twice as accurate as 1.0, with word error rate on a 19-language short-phrase set falling from 20.6% to 6.8%. These figures are vendor-reported.
• Default model: the documentation disagrees on which model is the default, so name the model explicitly.
• Deprecation: xAI says version 1.0 will be deprecated in the coming weeks.
• Lineage: the company says it draws on the audio foundation model behind Grok Voice.