---
license: apache-2.0
library_name: mlx
tags:
  - speech-enhancement
  - audio
  - mlx
---

# MossFormer2 SE 4-bit (MLX)

48kHz speech enhancement model converted to MLX format (4-bit quantized).

## Original Model
[alibabasglab/MossFormer2_SE_48K](https://huggingface.co/alibabasglab/MossFormer2_SE_48K)

## Usage

```python
from mlx_audio.sts.models.mossformer2_se import MossFormer2SEModel

model = MossFormer2SEModel.from_pretrained("starkdmi/MossFormer2-SE-4bit")
enhanced = model.enhance("noisy.wav")
```

## Precision Variants
- [MossFormer2-SE](https://huggingface.co/starkdmi/MossFormer2-SE) (fp32, 211MB)
- [MossFormer2-SE-fp16](https://huggingface.co/starkdmi/MossFormer2-SE-fp16) (fp16, 106MB)
- [MossFormer2-SE-8bit](https://huggingface.co/starkdmi/MossFormer2-SE-8bit) (int8, 86MB)
- [MossFormer2-SE-6bit](https://huggingface.co/starkdmi/MossFormer2-SE-6bit) (int6, 75MB)
- [MossFormer2-SE-4bit](https://huggingface.co/starkdmi/MossFormer2-SE-4bit) (int4, 64MB) ← this

## Performance
~30x real-time on Apple M4