SpaceXAI's Grok Voice Transcribe 2.0 cuts speech-to-text error rate to 6.8%

SpaceXAI has released Grok Voice Transcribe 2.0, an update to its speech-to-text system, software that turns spoken audio into text. The company says the new version is twice as accurate as Grok Voice Transcribe 1.0 while charging the same price. It runs only as a hosted API under the model ID grok-voice-transcribe-2.0; SpaceXAI has not released open weights, so it cannot be run on a company’s own servers.

The model is built on the audio foundation model that powers Grok Voice, which SpaceXAI says already handles tens of thousands of customer-support calls a day, transcribes millions of hours of video narration, and runs the Grok assistant inside Tesla vehicles. SpaceXAI trained Grok Voice Transcribe 2.0 on live, noisy, multilingual recordings gathered from those real-world environments, then refined it further with additional post-training.

SpaceXAI tested the model on four internal sets pulled from production traffic: telephone customer-support calls, conversations with Grok, spoken credentials such as phone numbers and addresses, and short voice-assistant phrases across 19 languages. Version 2.0 improved on version 1.0 across all four, and the company says it leads every model it tested on the telephony set. The largest gain came on short phrases, where word error rate, or WER, the share of words transcribed incorrectly, fell from 20.6% to 6.8%, a drop of roughly 67% in the number of word errors. SpaceXAI also reports that the model ranks first among 32 systems on the AA-WER Streaming benchmark run by Artificial Analysis, a public leaderboard for transcription accuracy. These results come from SpaceXAI itself and have not been independently reproduced.

Atlassian Loom adopts the model

Atlassian’s video tool Loom has switched to Grok Voice Transcribe 2.0 for transcribing recordings, after finding it more accurate than its previous transcription setup. Sanchan Saxena, Atlassian’s SVP of Teamwork Collection, described the change as “closing the loop from context to code.” In the workflow SpaceXAI describes, a user records a plan in Loom, the transcript is fed into the coding tool Cursor, and Cursor uses it to make the actual code changes.

Version 2.0 keeps the pricing of its predecessor: $0.10 per hour of audio for batch transcription and $0.20 per hour for real-time streaming, with speaker diarization, tagging which speaker said each line, word-level timestamps, and support for up to 100 custom key terms included at no extra cost. The API also removes filler words like “um” and “uh” by default and can transcribe up to eight audio channels separately in one request. Batch files can run up to 500 MB across 12 formats, and streaming supports compressed Opus audio to cut bandwidth use.

SpaceXAI has not said whether outside researchers will get access to verify its accuracy claims. Anyone integrating the API today needs to set the model parameter to grok-voice-transcribe-2.0 explicitly, since it is not yet the default.