SpaceXAI has released Grok Voice Transcribe 2.0, a speech-to-text model that reportedly reduces error rates by half compared to its predecessor while maintaining the same cost. The model is designed to handle complex audio, including background noise, overlapping speech, and regional dialects.
The model is built on the audio technology used in Grok Voice, which currently processes tens of thousands of support calls daily and powers Tesla's voice assistant. In the Artificial Analysis streaming benchmark, it ranked first among 32 models based on the AA-WER Streaming metric.
The update features enhanced multilingual support, allowing the system to automatically detect and switch between dozens of languages in real time. On short phrases, the Word Error Rate (WER) dropped from 20.6% to 6.8%. The API also supports automatic formatting for numbers, dates, and currencies across 25 languages.
The model is accessible via a Speech to Text API in both batch and streaming modes. Batch processing is priced at $0.10 per hour of audio, while streaming costs $0.20 per hour. The API supports up to eight audio channels, custom keyword lists, and speaker diarization at no extra charge.
