Voxtral hits browser with WebGPU transcription

This is less about one flashy demo and more about speech AI finally becoming a practical browser primitive.

• –Running Voxtral locally cuts latency, cloud cost, and privacy risk for captions, assistants, and meeting tools.
• –The underlying model targets sub-500 ms transcription across 13 languages, so this is showing real capability rather than a toy browser experiment.
• –Transformers.js plus WebGPU dramatically lowers the barrier for web teams that want on-device ASR without native wrappers.
• –The constraint is still browser and GPU support, but the broader trend is clear: open speech models are getting much closer to proprietary realtime APIs in usability.