reza2kn/vosk-rust

Vosk-Rust is a from-scratch rewrite, entirely in the Rust programming language, of Vosk, a popular speech-to-text system. Vosk normally depends on Kaldi, an older speech recognition toolkit written in C++, which makes it awkward to package and run on small devices. This project removes that C++ dependency entirely: you feed in a standard Vosk model folder and raw audio, and it prints out the recognized words, without needing libvosk, Kaldi, or Python installed. The README states plainly that on its test audio clip, this Rust version produces text that matches the original Vosk exactly. Each stage of the pipeline, reading the model files, converting audio into the numeric features the model expects, running the neural network, and decoding those outputs into words, was checked one by one against the original Kaldi based tools to confirm the math lines up. The author is honest about where the Rust version currently falls short of the original on harder audio. On difficult Persian speech samples, this rewrite has a small but consistent accuracy gap compared to the original, and the README traces the entire gap to one specific piece: the original feeds the model extra information about the speaker's voice and background noise, while this version currently feeds in blank placeholders for that information. When the README swapped in the real values just to test, the gap disappeared. On clean audio, the two versions perform about the same. The project was originally built to support a separate keyword-spotting tool for Persian speech, but the README notes it works with any compatible Kaldi-style acoustic model, not just that one. It includes an experimental feature to shrink model file size, which the author explicitly recommends against using because it noticeably hurts accuracy on harder audio, so the default full-size model is what ships. Vosk-Rust is released under the Apache-2.0 license, a permissive license allowing commercial and personal use.