xAI Launches Grok Voice Transcribe 2.0 With Double the Accuracy
xAI released Grok Voice Transcribe 2.0 on September 18, 2026, a speech-to-text model the company says is roughly twice as accurate as its predecessor at the same price, built on the audio foundation model behind Grok Voice.
Details
- Leaderboard position: ranks first for accuracy among 32 streaming transcription services on the public Artificial Analysis leaderboard
- Biggest gain is multilingual: error rates on short phrases dropped from 20.6% to 6.8% compared to version 1.0
- Built for messy real audio: trained on live, noisy, multilingual recordings captured across diverse environments and refined with post-training, targeting noisy phone calls, multiple speakers, and varied accents
- Feature set: batch and streaming transcription, word-level timestamps with confidence scores, speaker identification, multichannel processing, custom term biasing, intelligent formatting, filler removal, and turn detection for voice agents
- Pricing unchanged: batch transcription stays at $0.10 per audio hour, streaming at $0.20 per hour, with diarization, timestamps, and key-term biasing included at no extra charge
- Existing footprint: Grok Voice already handles tens of thousands of customer-support calls a day, transcribes millions of hours of video narration, and runs voice agents in products including the Grok assistant in Tesla vehicles
What happened next
Atlassian Loom adopted Grok Voice Transcribe 2.0, telling xAI it outperformed their previous transcription solution. The release follows closely on Grok Build’s new memory system from September 16, continuing xAI’s pattern of shipping incremental infrastructure upgrades across its Grok product line rather than a single flagship model update. xAI positioned the model as a drop-in upgrade for existing Voice Transcribe customers rather than a new standalone product.