Grok Voice Transcribe 2.0: 5 Details That Matter

xAI shipped Grok Voice Transcribe 2.0 today, and the headline number — twice the accuracy of its predecessor — undersells how specific the improvements actually are. From short voice commands to credentials like phone numbers and email addresses, the model targets the exact failure modes that made first-generation speech transcription frustrating in real products. Here are the five details worth paying attention to.

1. The accuracy gains are specific, not vague

xAI isn't just claiming a broad improvement — the benchmarks are granular. According to xAI, the word error rate for short phrases dropped from 20.6% to 6.8%. For conversational audio, it fell from 8.7% to 3.3%. Credential transcription (phone numbers, email addresses) improved from a 7.2% error rate to 3.2%. In telephony tests using compressed 8 kHz audio — the kind of signal quality you get on a standard phone call — error rates dropped from 10.6% to 7.1%. These aren't cherry-picked demo conditions; telephony compression and background noise are the hardest real-world scenarios for speech recognition, and Transcribe 2.0 was specifically built to handle them.

2. It ranks first on the public Artificial Analysis leaderboard

According to xAI, Transcribe 2.0 currently sits at the top of the Artificial Analysis leaderboard for accuracy across 32 streaming models. That's a publicly verifiable benchmark, not an internal claim — which gives developers a meaningful reference point when evaluating whether to switch from existing solutions. First place on a 32-model leaderboard is a strong position to launch from, though leaderboard rankings shift as competitors update their own models.

3. Pricing is unchanged from version 1.0

Transcribe 2.0 is available now through the Grok Voice API at $0.10 per hour for batch processing and $0.20 per hour for streaming — the same rates as the first version. That means any developer or business already using Transcribe 1.0 gets a substantial accuracy upgrade at zero additional cost. Speaker diarization, word-level timestamps, and key term biasing are also included at no extra charge, according to xAI. For teams running high-volume transcription workloads, the batch tier in particular represents solid value against the accuracy improvements on offer.

▶ Watch Video on X

4. Atlassian is already shipping it inside Loom

The most concrete third-party validation at launch comes from Atlassian, which integrated Transcribe 2.0 into Loom and confirmed it outperforms their previous solution for capturing user instructions across all videos. Sanchan Saxena, SVP of Teamwork Collection at Atlassian, described the integration as helping to "close the loop from context to code" — specifically calling out a workflow where users dictate change requests in Loom, then export directly to Cursor for coding. That's a real production deployment, not a pilot, and it signals that enterprise-grade reliability is already there on day one.

5. The Tesla connection is real — and worth watching

For Tesla owners, Grok Voice Transcribe 2.0 matters beyond the developer API. Tesla's in-car voice assistant and FSD-related voice command processing sit in the same technology ecosystem as xAI's models — and accuracy improvements in short voice commands and noisy environments map directly onto the scenarios where in-car voice recognition most often fails. The word error rate on short phrases dropping from 20.6% to 6.8% is the kind of improvement that closes the gap between a voice command that works reliably and one that requires three attempts. Whether and when xAI's transcription improvements reach Tesla's vehicle software hasn't been announced, but the trajectory is clear.

Transcribe 2.0 is available in the Grok Voice API as of today. Developers can find full documentation and the xAI blog post linked in the source tweets above.

This report was curated by the BASENOR Editorial Desk from the sources listed above. Read our editorial standards or email editorial@basenor.com to report an error.