xAI Releases Grok Voice Transcribe 2.0: 2x Accuracy Improvement, Price Held Steady

xAI has released its latest speech recognition model, Grok Voice Transcribe 2.0. This model doubles transcription accuracy compared to the previous 1.0 version while keeping the price unchanged. It ranks first in accuracy among 32 streaming models on the Artificial Analysis public leaderboard and is built on an audio foundation model validated in real-world services such as the in-car voice assistant in Tesla vehicles.

Performance and Benchmarks

According to internal evaluation data, performance improved across all four categories, including customer support calls and multilingual commands. Notably, the Word Error Rate (WER) for short in-car commands with limited context dropped significantly from 20.6% to 6.8%. In 8kHz telephone call environments, it achieved the highest performance among all tested models, with particularly notable improvements in multilingual transcription accuracy.

Key Features and Compatibility

It is compatible with the existing Speech-to-Text API, allowing immediate adoption without code changes. Key features include Batch and Streaming support, speaker diarization, independent processing of up to 8 channels, and key term biasing. It also provides filler word removal (e.g., 'um', 'uh') and speaker turn detection for voice agents.

Pricing and Adoption Cases

Pricing is set at $0.10 per hour for Batch transcription and $0.20 per hour for Streaming transcription, with speaker diarization and timestamping included by default. Atlassian Loom adopted Grok Voice Transcribe 2.0 for video transcription, judging it to be more accurate than existing solutions, and built a workflow that passes transcripts to Cursor to automate code updates. Grok Voice Transcribe 2.0 will soon become the API default, with support for version 1.0 scheduled to end within a few weeks.