On September 18 local time, SpaceX's artificial intelligence department, SpaceXAI, released its new speech-to-text model Grok Voice Transcribe2.0. The most impressive feature is that it reduced the error rate by about half without increasing the price.

It is based on the underlying audio foundation model of Grok Voice, which has now been deeply integrated into real business operations: handling tens of thousands of customer service calls daily, transcribing millions of hours of video voiceovers, and driving various voice assistants in physical hardware, including the Grok assistant in Tesla vehicles. In other words, it is not a model that just runs benchmarks in the lab, but a product that has been repeatedly refined by massive real-world voice data.

image.png

On the public ranking list of Artificial Analysis, Grok Voice Transcribe2.0 ranked first among all 32 streaming models, leaving its competitors far behind. SpaceXAI is not satisfied with public benchmarks; it sampled four internal datasets from its own online services for systematic evaluation, covering customer service call recordings, daily English conversations with Grok, structured information such as spoken phone numbers and email addresses, and multilingual short instructions—on all four datasets, version 2.0 completely outperformed version 1.0.

image.png