xAI has released Grok Voice Transcribe 2.0, a speech-to-text model the company says is twice as accurate as its prior version, Grok Voice Transcribe 1.0, at unchanged pricing: $0.10 per audio hour for batch jobs, $0.20 per hour for streaming. That accuracy claim comes from xAI’s own real-world evaluations, not an independent audit.
The model runs on the same audio foundation as Grok Voice, which xAI says handles a high daily volume of customer-support calls, does the narration transcription for millions of hours of video, and powers the voice assistant built into Tesla vehicles. That existing footprint is the pitch: xAI is framing transcription accuracy as a byproduct of a voice product already deployed at scale, not a standalone research exercise.
Most transcription models handle clean audio well, xAI argues, and stumble once conditions get messy: a caller reading out an account number, a staticky or dropped line, two speakers talking over one another, a thick regional accent. Grok Voice Transcribe 2.0 was built specifically for that harder case, trained on live, noisy, multilingual recordings gathered across varied environments and then sharpened with additional post-training.
On the public Artificial Analysis leaderboard, xAI says the model ranks first for accuracy among 32 streaming transcription systems. That leaderboard is third party, though the rest of the accuracy story rests on xAI’s own internal testing. xAI checked the new model against four categories drawn from real production traffic: short multilingual voice commands, calls where a speaker reads out an account code or similar credential, ordinary phone support calls, and recorded conversations people have with Grok itself. Grok Voice Transcribe 2.0 beat its predecessor across all four categories, and on the phone support calls specifically it outperformed every other model xAI tested.
Multilingual handling is where xAI says the improvement is largest. Grok Voice Transcribe 2.0 detects language automatically, follows switches mid-recording, and covers dozens of languages in one pass. On short phrases, the kind of clipped command a driver might give a car, xAI reports the word error rate falling from 20.6 percent to 6.8 percent.
The release also adds features aimed at developers building on the API: word-level timestamps with confidence scores, free speaker diarization, transcription across up to eight audio channels at once, and the ability to bias recognition toward up to 100 custom terms per request, useful for product names or clinical vocabulary. It also adds automatic formatting of numbers and contact details, filler-word removal, and turn detection tuned for voice agents. Existing integrations get the upgrade with no code changes once xAI flips the default, though teams that need to stay on the old model can pin grok-voice-transcribe-1.0 before xAI retires it in the coming weeks.
The clearest evidence of real-world traction is Atlassian’s Loom, the screen-recording tool. xAI says Loom found Grok Voice Transcribe 2.0 more accurate than its previous transcription vendor. “We’ve always believed the best way to move work forward is to capture context once and let it flow everywhere,” said Sanchan Saxena, Atlassian’s senior vice president of Teamwork Collection, in comments shared by xAI. He described a workflow where a recorded Loom walkthrough feeds directly into the Cursor coding tool, turning a spoken plan into code changes.
That Loom integration is the more interesting story than the benchmark table. It points to xAI selling transcription less as a commodity API and more as the connective layer between voice, video, and coding agents, a bet that speech becomes the input format for software work rather than typed instructions. Speech-to-text is also a crowded market with entrenched incumbents, and xAI has not published a direct comparison of its pricing or accuracy against rivals outside its own leaderboard citation.
Teams already piping call or meeting transcripts into downstream AI agents should benchmark Grok Voice Transcribe 2.0 against their current provider before the automatic default switch in the coming weeks.
Based on xAI’s official announcement, “Introducing Grok Voice Transcribe 2.0,” published on x.ai.
