xAI Says Grok Voice Transcribe 2.0 Doubles Accuracy at Same Price

Grok Voice Transcribe 2.0 is now available from xAI, with the company saying the new speech-to-text model is twice as accurate as Grok Voice Transcribe 1.0 at the same price. The model is built for real-world audio, including noisy phone calls, overlapping speakers, accents, spoken credentials, and short multilingual commands.

The release matters beyond basic dictation. Grok Voice Transcribe 2.0 supports both batch and real-time streaming transcription, speaker diarization, word-level timestamps, multichannel audio, domain-specific key terms, formatting for numbers and contact details, filler-word removal, and turn detection for voice agents. Existing Speech-to-Text API integrations get the new accuracy improvement without code changes, according to xAI.

Grok Voice Transcribe 2.0 Targets Real-World Audio

Grok Voice Transcribe 2.0 was trained on live, noisy, multilingual audio from different environments. xAI says the model is designed for conditions where clean benchmark recordings are not enough, such as weak phone connections, competing voices, local accents, and phone numbers or email addresses spoken aloud.

On the public Artificial Analysis leaderboard, xAI says Grok Voice Transcribe 2.0 ranks first for accuracy among 32 streaming speech-to-text models. xAI also reports improvements over version 1.0 across four internal evaluation sets covering customer-support calls, conversations with Grok, spoken credentials, and short multilingual voice commands.

For short multilingual phrases, xAI reports that word error rate fell from 20.6% with version 1.0 to 6.8% with version 2.0. That is an internal evaluation result, so it should not be treated as a general benchmark result across every audio workload.

Multilingual Transcription Gets a Large Upgrade

xAI says multilingual accuracy is the biggest improvement over the previous model. Grok Voice Transcribe 2.0 can automatically detect languages and follow a language switch during the same recording.

The API documentation lists formatting support for languages including English, Arabic, Hindi, Japanese, Korean, Spanish, French, German, Portuguese, Turkish, Indonesian, Malay, Thai, Vietnamese, and others. The model can transcribe speech across the supported language set even when the language-formatting parameter is not used.

The API Adds More Control for Developers

• Batch transcription for recorded audio files and URLs.
• Streaming transcription over WebSocket for real-time applications.
• Word-level timestamps with confidence scores.
• Speaker diarization at no additional charge.
• Multichannel transcription for up to eight channels.
• Up to 100 key terms per request for names, products, or specialist vocabulary.
• Automatic formatting for numbers, dates, currencies, phone numbers, and email addresses.
• Optional filler-word removal.
• Smart Turn detection for voice-agent conversations.

The feature set makes the model useful for more than meeting transcripts. Developers can use streaming for voice applications, multichannel transcription for recordings where speakers are separated by audio channels, and key-term biasing when a system needs to recognize product names or specialized vocabulary.

Loom Is Already Using It for Video Transcription

Atlassian says Loom now uses Grok Voice Transcribe 2.0 to transcribe every video after finding it more accurate than its previous solution. xAI highlights a workflow in which a Loom recording captures an action plan, the transcript is passed into Cursor, and the resulting instructions can drive code changes.

That use case connects speech recognition to the broader agent workflow already appearing across Saganote. Grok Build Mode lets SuperGrok Heavy subscribers create and host apps from chat prompts, while Grok for Excel brings Grok into Microsoft 365 spreadsheets.

Pricing Stays the Same

xAI is keeping Grok Voice Transcribe 2.0 at the same rates as version 1.0: $0.10 per hour for batch transcription and $0.20 per hour for streaming. xAI says diarization, timestamps, and key terms are included at those prices.

The model is available now through xAI’s Speech-to-Text API. The developer documentation identifies grok-voice-transcribe-2.0 as the model name.

Version 2.0 Will Replace the First Model

xAI says Grok Voice Transcribe 2.0 will soon become the default model in its Speech-to-Text API, while Grok Voice Transcribe 1.0 will be deprecated in the coming weeks. Developers that need to stay on version 1.0 during the transition can pin grok-voice-transcribe-1.0.

For developers already using xAI’s Speech-to-Text API, the release is primarily an accuracy and capability upgrade rather than a pricing change. The biggest reported gains are in multilingual audio and difficult real-world recordings, while the existing API integration can continue without code changes.

Source: xAI announcement and xAI Speech-to-Text documentation.