SpaceXAI has updated its speech-to-text capabilities with the launch of Grok Voice Transcribe 2.0. Available through the Grok Voice API, the model supports both batch and streaming audio processing. The company reports that this version provides a significant leap in performance over its predecessor, specifically noting a twofold increase in accuracy while maintaining the previous pricing structure.
Performance and Pricing
The most notable technical improvement involves the word error rate (WER) for short-phrase transcriptions. According to figures reported by MarkTechPost, the WER across a set of 19 languages has dropped from 20.6% in version 1.0 to 6.8% in version 2.0. This reduction suggests that the new model is substantially more reliable for brief audio clips, which are often challenging for automated transcription services. Despite these technical gains, SpaceXAI has kept the cost of the service stable. Batch processing remains priced at $0.10 per hour, while real-time streaming audio transcription costs $0.20 per hour.
Technical Specifications and Language Support
The announcement did not specify the exact architecture or training methodology used to achieve these accuracy gains. While the improvement in short-phrase WER is documented across 19 languages, the specific list of supported languages was not provided in the initial release. Furthermore, it remains unclear how the model performs on long-form audio or in environments with significant background noise, as the reported data focused on short phrases. SpaceXAI has also not detailed any changes to the API’s latency or throughput limits. The company has not stated if version 1.0 will be deprecated or if users can continue to choose between models for specific workloads.
Developers can access the updated model immediately through the existing API infrastructure.
Lead image: Egor Komarov / Pexels
