Voxtral Mini 4B Realtime WebGPU brings free live transcription to browsers

Voxtral Mini 4B Realtime WebGPU is distributed through Hugging Face, where the model page lets users start transcription with a single click. On first use, the browser downloads about 2.8GB of model data; later visits can reuse the cached copy. The interface is minimal: a microphone button starts listening, and recognized text appears in a Transcript pane almost in real time.

The model handles 13 languages — Japanese, English, Chinese, Korean, French, German, Spanish, Italian, Portuguese, Dutch, Russian, Arabic, and Hindi — without a language-selection menu. Instead, it detects the spoken language automatically. It is built purely for speech-to-text; translation, speaker labeling, and audio-file upload are not included.

Processing happens entirely on the local machine. The AI model is loaded from Hugging Face's Transformers.js as an ONNX-format model and runs on the PC's GPU via WebGPU, so audio does not need to be sent to a remote server. That local approach also sets the hardware bar: the browser and GPU must support WebGPU and shader-f16 (16-bit floating-point calculations). One report notes an error on a PC with 4GB of dedicated VRAM, and during a test with 8GB of GPU VRAM, memory use rose from 2.0GB before launching the tool to 5.9GB while it ran.