Pure C Implementation of Mistral Voxtral Realtime 4B Speech-to-Text Model Inference

Summary

This article describes a pure C implementation of the inference pipeline for Mistral AI's Voxtral Realtime 4B speech-to-text model. The implementation has zero external dependencies beyond the C standard library and features audio processing with chunked encoding and overlapping windows to bound memory usage regardless of input length. It supports audio input from stdin, microphone capture (macOS), and includes MPS inference with decent speed, though BLAS acceleration is described as usable but slow due to continuous bf16 to fp32 weight conversion.