QwenASR

A fast, pure-Rust, CPU-only inference engine for Qwen3-ASR speech-to-text, tuned for Apple Silicon.

Tech stack	Pure Rust, CPU-only — no Python, no GPU, no tensor framework, only libc
Optimization	Hand-written NEON / Accelerate / AMX-aware kernels for Apple Silicon
Performance	Transcribes a 28-second clip in 613 ms on an M5 Pro (46× realtime)
Features	Offline & streaming transcription, live capture with VAD, SRT/VTT subtitles, structured JSON, forced alignment
Install	cargo install qwen-asr-cli
Website	github.com/huanglizhuo/QwenASR

QwenASR is a fast, pure-Rust, CPU-only inference engine for Qwen3-ASR speech-to-text, tuned for Apple Silicon with hand-written NEON / Accelerate / AMX-aware kernels.

Performance

It transcribes a 28-second clip in 613 ms on an M5 Pro (46× realtime), beating GPU-based MLX implementations.

Features

Offline and streaming transcription
Live capture with VAD
SRT/VTT subtitles
Structured JSON output
Forced alignment

Install

cargo install qwen-asr-cli

FAQ

Do I need a GPU or Python to run QwenASR?

No. It is a pure-Rust, CPU-only inference engine — no Python, no GPU, no tensor framework, only libc.

How fast is it?

It transcribes a 28-second clip in 613 ms on an M5 Pro (46× realtime), beating GPU-based MLX implementations.

Visit project →