Skip to main content

Grok Voice Transcribe 2.0 Cuts Error Rate in Half, Keeps Price Steady

SpaceXAI's New Speech Model: Half the Errors, Same Price

On September 18, SpaceX's AI division, SpaceXAI, unveiled Grok Voice Transcribe 2.0—a speech-to-text model that cuts error rates by roughly 50% while keeping the price tag unchanged. That's a rare combo in the AI world, where better usually means pricier.

But this isn't just a lab experiment. The model is built on Grok Voice's underlying audio foundation, which already powers real-world operations: handling tens of thousands of customer service calls daily, transcribing millions of hours of video voiceovers, and driving voice assistants in physical hardware—including the Grok assistant inside Tesla vehicles. In other words, it's been battle-tested by massive amounts of actual voice data, not just benchmark tests.

Image

Topping the Charts

On Artificial Analysis's public leaderboard, Grok Voice Transcribe 2.0 ranked first among all 32 streaming models, leaving competitors in the dust. But SpaceXAI didn't stop at public benchmarks. They pulled four internal datasets from their own live services—customer service recordings, everyday English chats with Grok, structured info like spoken phone numbers and email addresses, and multilingual short commands. Across all four, version 2.0 beat version 1.0 handily.

Image

So what does this mean for you? If you're using voice tech for business or everyday tasks, you're getting significantly better accuracy without paying more. And with Tesla and other hardware already on board, this upgrade could quietly improve your daily interactions with voice assistants.

Key Points

  • Error rate halved: Grok Voice Transcribe 2.0 reduces errors by about 50% compared to its predecessor.
  • Price unchanged: No increase in cost despite the performance boost.
  • Real-world tested: Already handling customer calls, video voiceovers, and Tesla voice assistants.
  • Leaderboard champ: Ranked #1 among 32 streaming models on Artificial Analysis.
  • Internal validation: Outperformed version 1.0 on four internal datasets, covering diverse speech scenarios.