Voxtral Mini 4B Realtime (Rust)
A pure Rust implementation of Mistral's Voxtral Mini 4B speech recognition and text-to-speech models, running natively or in the browser through the Burn machine-learning framework.
voxtral-mini-realtime-rs reimplements Mistral's Voxtral Mini 4B Realtime automatic speech recognition and text-to-speech models in Rust using the Burn machine-learning framework, running either as a native application or in the browser through WebAssembly and WebGPU. It supports both full-precision BF16 weights and a quantized Q4 GGUF path with a custom WGSL shader that fuses dequantization with matrix multiplication for the browser build.
On an NVIDIA DGX Spark, the Q4 GGUF path reaches a real-time factor of about 0.42 for speech recognition on a 16-second test clip, faster than the BF16 path, with a measured 8.49% word error rate on the FLEURS English benchmark against Mistral's own reported 4.90% at full precision. The project also documents and works around a decoder padding issue that makes the quantized model produce blank output on audio clips with no leading silence.