| --- |
| license: apache-2.0 |
| tags: |
| - mlx |
| - speech-to-text |
| - speech |
| - transcription |
| - asr |
| - stt |
| - mlx-audio |
| library_name: mlx-audio |
| pipeline_tag: automatic-speech-recognition |
| base_model: |
| - zhifeixie/Mega-ASR |
| --- |
| |
| # Mega-ASR MLX Q4 |
|
|
| This is a private MLX conversion of `zhifeixie/Mega-ASR`. |
|
|
| The checkpoint was produced by merging the `mega-asr-merged` LoRA adapter from |
| `zhifeixie/Mega-ASR` into the bundled `Qwen3-ASR-1.7B` base checkpoint, then |
| converting the merged weights to the `mlx-audio` Qwen3-ASR layout. |
|
|
| ## Conversion |
|
|
| - Base/source repo: `zhifeixie/Mega-ASR` |
| - Adapter: `mega-asr-merged` |
| - Format: MLX / `mlx-audio` |
| - Quantization: affine Q4, `group_size=64`, `bits=4` |
| - Text model and token embedding are quantized; audio tower remains full precision. |
|
|
| ## Use With mlx-audio |
|
|
| ```bash |
| pip install -U mlx-audio |
| python -m mlx_audio.stt.generate --model runfuture/Mega-ASR-MLX-Q4 --audio audio.wav |
| ``` |
