xAI's new transcription model doubles accuracy at no extra cost

Speech-to-text models usually breeze through clean studio recordings, but throw in a patchy cell connection, screaming traffic, or someone rattling off an email address, and most engines fall apart fast.

xAI just dropped Grok Voice Transcribe 2.0, taking aim squarely at those nightmare audio conditions. According to the company, the upgraded model doubles the accuracy of its predecessor while keeping prices flat, and it just secured the number one accuracy spot among 32 streaming models on the Artificial Analysis benchmark.

Built for flaky phone lines and chaotic accents

The new engine builds on the foundation model behind Grok Voice, which already processes tens of thousands of customer support calls daily, transcribes millions of hours of video, and runs voice commands inside Tesla vehicles.

Instead of training purely on pristine datasets, xAI trained and post-trained this model on messy, multilingual, real-world field recordings. In internal evaluations across customer service calls, Grok voice chats, spoken credentials, and voice-assistant commands, the 2.0 version outperformed version 1.0 across the board, topping every tested competitor on telephony audio.

Multilingual recognition saw the most dramatic leap. The model detects languages automatically on the fly and seamlessly tracks mid-sentence language switches.

On a benchmark of short voice commands across 19 languages—notoriously hard because brief utterances offer little context—the word error rate plummeted from 20.6% down to 6.8%.

Same price, immediate drop-in replacement

Beyond raw benchmarks, enterprise adoption is already rolling out. Atlassian adopted the model to transcribe every screen capture on Loom, pointing to workflows where users talk through feature plans on video and dump the resulting transcript straight into Cursor to generate code automatically.

The pricing strategy is aggressive. Batch transcription stays at $0.10 per audio hour, while live streaming runs $0.20 per audio hour. Both tiers bundle speaker diarization, timestamps, and custom key terms at no extra cost.

Existing API users do not need to rewrite their integrations to get the upgrade. The 2.0 version will soon become the default endpoint, and Grok Voice Transcribe 1.0 is scheduled for deprecation in the coming weeks.

Why it matters

Speech-to-text has felt commoditized lately, with many providers treating audio transcription as an afterthought bundled into broader multimodal stacks.

We think xAI is making a clever, highly practical move here. Rather than fighting for marginal gains on clean audio benchmarks, they weaponized the massive, noisy acoustic data coming out of Tesla cabins and real-world support desks to dominate dirty audio conditions.

For builders, at ten to twenty cents an hour with native diarization included, this makes spinning up automated call center QA, meeting summaries, and voice agents substantially cheaper and far more reliable in the field. If your current pipeline struggles with muffled voices or mixed languages, it is well worth testing today.

The main friction point will be platform lock-in and enterprise trust around data routing, but in terms of price-to-performance on noisy speech, xAI just put serious pressure on Whisper and legacy transcription APIs.