xAI launches Grok Voice Transcribe 2.0 for real-world audio

xAI has launched Grok Voice Transcribe 2.0, a speech-to-text model that, according to the company’s tests, doubles the accuracy of its previous version without increasing the price. It is designed for noisy calls, multiple speakers, local accents, and difficult data such as phone numbers and email addresses.

The improvement is not limited to clean recordings of one person speaking into a microphone. The model was trained on real-world, multilingual audio from different environments, including those already used by Grok’s voice systems.

More accurate in difficult situations

In the Artificial Analysis public ranking, Grok Voice Transcribe 2.0 ranks first in accuracy among 32 streaming transcription models. xAI also compared it with the previous version using four internal datasets based on production traffic:

• Customer service calls in English.
• Conversations with Grok.
• Phone numbers, email addresses, and addresses spoken aloud.
• Voice commands in 19 languages.

In all of them, the new version improves on Grok Voice Transcribe 1.0. For phone calls, xAI says it outperforms every model it tested.

The biggest improvement is in language handling. The model automatically detects which language is being spoken and can follow a language switch during the same recording. In short phrases, where there is little context to identify the language, the word error rate drops from 20.6% to 6.8%.

That detail matters for voice assistants. A command such as "turn on the lights" or an email address contains few words, and any error can make the result unusable.

What changes for existing users

Existing integrations with the speech-to-text API receive the improvement without code changes. Atlassian, for example, tested the model in Loom, its service for recording and sharing screen videos, and found it more accurate than its previous solution.

More reliable transcription also makes it possible to automate follow-up tasks. A user can record a work plan in Loom, send the text to Cursor, and use it as the basis for making code changes.

The price stays the same:

• Batch transcription: $0.10 per hour of audio.
• Streaming transcription: $0.20 per hour.
• Diarization, timestamps, and key terms included.

Diarization identifies which person is speaking at each moment. It is especially useful in meetings and support calls.

The previous version will be retired

xAI plans to make Grok Voice Transcribe 2.0 the default model for its speech-to-text API. Grok Voice Transcribe 1.0 will be retired in the coming weeks, according to the announcement.

Applications that need to keep the older version will have to explicitly set the identifier grok-voice-transcribe-1.0. For everyone else, the change will happen automatically when the new version becomes the default option.

The important signal is not just that the model makes fewer errors in a test. It is that voice transcription is starting to focus on the audio that actually exists: imperfect calls, overlapping conversations, language switches, and data that must be written down exactly. That is where it will be decided whether voice assistants can carry out useful tasks without constant human correction.