Post
738
Qwen3-ASR Swift: On-Device Speech Recognition for Apple Silicon
I'm excited to release https://github.com/ivan-digital/qwen3-asr-swift, an open-source Swift implementation of Alibaba's
Qwen3-ASR, optimized for Apple Silicon using MLX.
Why Qwen3-ASR? Exceptional noise robustness — 3.5x better than Whisper in noisy conditions (17.9% vs 63% CER).
Features:
- 52 languages (30 major + 22 Chinese dialects)
- ~600MB model (4-bit quantized)
- ~100ms latency on M-series chips
- Fully local, no cloud API
https://github.com/ivan-digital/qwen3-asr-swift | Apache 2.0
I'm excited to release https://github.com/ivan-digital/qwen3-asr-swift, an open-source Swift implementation of Alibaba's
Qwen3-ASR, optimized for Apple Silicon using MLX.
Why Qwen3-ASR? Exceptional noise robustness — 3.5x better than Whisper in noisy conditions (17.9% vs 63% CER).
Features:
- 52 languages (30 major + 22 Chinese dialects)
- ~600MB model (4-bit quantized)
- ~100ms latency on M-series chips
- Fully local, no cloud API
https://github.com/ivan-digital/qwen3-asr-swift | Apache 2.0
