xAI released Grok Voice Transcribe 2.0 on September 18, a new speech-to-text model the company says is twice as accurate as version 1.0 while costing exactly the same. According to xAI, the model tops the Artificial Analysis leaderboard for accuracy among 32 streaming transcription models, and it’s already running inside Atlassian Loom.
The pitch is simple: better transcripts for hard audio, no price bump, and no code changes for anyone already on the API. Here’s what actually shipped.
It’s Built for Messy Audio, Not Studio Recordings
xAI is upfront that most transcription models do fine on clean, single-speaker audio. The hard part is the real world: bad phone lines, people talking over each other, regional accents, and someone reading an email address out loud. Grok Voice Transcribe 2.0 is trained on what xAI calls “live, noisy, multilingual audio recorded across a diverse set of environments.”
The model shares its audio foundation with Grok Voice, which xAI reports already handles tens of thousands of customer-support calls a day and runs the Grok assistant inside Tesla vehicles. So this isn’t a lab model getting its first taste of production traffic.
The Benchmarks Focus on Production-Style Tests
Beyond the public leaderboard, xAI ran four internal word-error-rate tests pulled from actual traffic:
- Telephony: 8 kHz customer-support calls in English. xAI says 2.0 beats every model it tested here.
- Conversational: English conversations with Grok.
- Credentials: phone numbers, emails, and addresses read aloud.
- Short phrases: voice-assistant commands across 19 languages.
Version 2.0 improves on 1.0 across all four. The standout number is the short-phrase set, where word error rate dropped from 20.6% to 6.8%. That’s the in-car command scenario where the model gets almost no context to work with.
Multilingual Is the Biggest Jump
xAI calls multilingual accuracy the largest single improvement over 1.0. The model transcribes dozens of languages, detects the language on its own, and follows mid-recording switches in a single pass. That last part matters for support centers and any product with bilingual users who flip between languages mid-sentence. xAI’s own chart puts 2.0 ahead of ElevenLabs Scribe v2 and Deepgram Nova-3 on multilingual word error rate, though that’s xAI’s internal comparison, so treat it accordingly.
The Feature List Covers Most Real-World Needs
All of this comes through the existing Speech-to-Text API:
- Batch and streaming: recorded files and URLs, or live audio streams.
- Word-level timestamps with start/end times and confidence scores.
- Speaker diarization at no extra cost.
- Multichannel transcription, up to 8 channels handled independently.
- Key term biasing: pass up to 100 domain terms per request (product names, medical vocabulary).
- Text formatting: numbers, dates, currencies, phone numbers, and emails come back in written form.
- Filler word removal for cleaner transcripts.
- Smart turn detection so voice agents know when a speaker is done.
What stands out here is diarization being free. Several competitors bill it as an add-on, and it’s one of the features people actually use.
Atlassian Loom Is the Launch Customer
Atlassian tested 2.0 against its existing transcription setup for Loom videos and found it more accurate, per xAI. The workflow they describe is telling: record an action plan in Loom, pipe the transcript into Cursor, and let it make the code changes. Sanchan Saxena, Atlassian’s SVP of Teamwork Collection, called it “a glimpse of where AI-assisted development is headed.”
This is the real story for anyone building products. Transcription accuracy used to be a nice-to-have. Once transcripts feed directly into agents that take action, every error becomes a wrong action downstream.
Pricing and Availability
Pricing is unchanged from 1.0:
- Batch: $0.10 per hour of audio
- Streaming: $0.20 per hour of audio
Diarization, timestamps, and key terms are included. The model is live now through the API and a browser demo, with docs available.
One caveat: 2.0 will soon become the default in the Speech-to-Text API, and xAI plans to deprecate 1.0 “in the coming weeks.” If you need the old behavior during migration, pin grok-voice-transcribe-1.0 explicitly. Otherwise your integration flips over automatically.
Why This Matters
Speech-to-text has turned into a price war, and xAI is competing on accuracy-per-dollar rather than raw cost. At $0.10 per hour for batch, that’s $6 per 1,000 minutes, and the leaderboard chart shows most rivals sitting at higher prices with worse error rates. If those numbers hold up in independent testing, this puts pressure on Deepgram, ElevenLabs, and the hyperscalers to either drop prices or improve their noisy-audio performance.
The deprecation timeline is worth watching. Anyone with production traffic on 1.0 has a few weeks to validate 2.0 on their own audio before the switch happens for them. Full benchmarks and API details are in xAI’s announcement.