Grok Voice Transcribe 2.0 launched on September 18 for difficult speech such as noisy calls, overlapping voices and account details read aloud. The developers report fewer errors than version 1.0 across four internal evaluations.
Those tests separately cover English customer-support calls, conversations with Grok, spoken credentials and short voice commands in 19 languages. On the short-phrase set, the reported word error rate fell from 20.6% to 6.8%. That measures this test's transcription errors, not a universal accuracy rate for every language or recording.
Pricing remains $0.10 per hour for batch transcription and $0.20 per hour for streaming. Speaker labels, timestamps and key-term hints are included. The release also supports independently transcribing up to eight audio channels.
At announcement, Grok Voice Transcribe 2.0 was due to become the API default, with version 1.0 scheduled for deprecation in the following weeks. Existing users could pin version 1.0 during the transition.