Project
Qwen ASR RS
A Rust port of Qwen3-ASR for CPU speech-to-text.
Qwen ASR RS is a Rust port of antirez/qwen-asr, a CPU inference engine for Qwen3-ASR speech-to-text models. It uses candle for the neural network work and keeps the original's CPU-first design.
Pipeline§
A WAV file is converted to a log-mel spectrogram, passed through the audio encoder, combined with prompt embeddings, and decoded into text tokens. The crate supports Qwen3-ASR-0.6B and Qwen3-ASR-1.7B, detecting the model variant from the model directory.
The command-line tool transcribes a WAV file and prints text to stdout with timing on stderr. A separate benchmark binary measures the full pipeline or the encoder alone. The crate is also usable as a library through Pipeline::load, transcribe, and transcribe_timed.
Where it sits§
This is a standalone Rust implementation rather than a wrapper around the Python model. Model weights come from Hugging Face, and the repository includes download and build scripts for the supported models.