---
title: OmniCodec
canonical_url: "https://www.modelscope.cn/models/ASLP-lab/OmniCodec"
md_url: "https://www.modelscope.cn/models/ASLP-lab/OmniCodec.md"
repository: ASLP-lab/OmniCodec
last_updated: 2026-04-09
license: apache-2.0
pipeline_tag: feature-extraction
tasks:
  - feature-extraction
parameters: 647.9M
tensor_type:
  - BF16
library_name:
  - safetensors
  - pytorch
frameworks:
  - pytorch
downloads: 13
stars: 0
---

# OmniCodec

> OmniCodec - ASLP-lab 在 ModelScope 开源的模型。OmniCodec: Low Frame Rate Universal Audio Codec with Semantic–Acoustic Disentanglement

ASLP-lab/OmniCodec 是 ModelScope 魔搭社区上的 647.9M 参数feature-extraction模型，采用 apache-2.0 许可。

- **Repository**: ASLP-lab/OmniCodec
- **License**: apache-2.0
- **Tasks**: feature-extraction
- **Parameters**: 647.9M
- **Downloads**: 13
- **Stars**: 0
- **Last updated**: 2026-04-09

Source: https://www.modelscope.cn/models/ASLP-lab/OmniCodec

---

# OmniCodec

OmniCodec: Low Frame Rate Universal Audio Codec with Semantic–Acoustic Disentanglement

- Demo Page: [OmniCodec Demo Page](https://hujingbin1.github.io/OmniCodec-Demo-Page/)
- Huggingface: [Huggingface](https://huggingface.co/ASLP-lab/OmniCodec)
- Arxiv: [Arxiv](https://arxiv.org/html/2603.20638v1)


## Overview

This repo contains:

- **Training**: `train.py` (Accelerate + GAN / WavLM-related losses per config)
- **Dataset**: `dataset.py` (multi-domain mixing; loads audio paths from `scp`)
- **Inference**: `infer.py` (reconstructs audio with a pretrained checkpoint)
- **Config**: `config/config_omnicodec.yaml`

## Environment

### Requirements

Install python dependencies:

```bash
pip install -r requirements.txt
```

Notes:

- `requirements.txt` contains an editable install line `-e OmniCodec/transformers-main`. Make sure the referenced path exists in your environment, or adjust/remove that line if you already have `transformers` installed.

## Data preparation (scp)

The training config expects 3 **scp files** (one per domain): speech / music / sound.

Each line in scp can be either:

- `utt_id /abs/or/rel/path/to/audio.wav`
- `/abs/or/rel/path/to/audio.wav` (utt will be inferred from filename)

Example:

```text
utt0001 /data/speech/utt0001.wav
utt0002 /data/speech/utt0002.wav
```

### What dataset does

For each item, `dataset.py` will:

- load audio with `librosa.load(..., sr=sample_rate, mono=True)`
- apply `librosa.util.normalize(wav) * 0.95`
- crop/pad/repeat to `segment_size` (default: 240000 samples @ 24kHz = 10s)
- return a dict: `{"wav": Tensor[T], "utt": str, "text": None}`

Failed samples return `None` and are filtered by `collate_fn` in `train.py`.

## Configure

Edit `config/config_omnicodec.yaml`:

- **Data**
  - `data.speech_train_shards_dir`: path to `speech.scp`
  - `data.music_train_shards_dir`: path to `music.scp`
  - `data.sound_train_shards_dir`: path to `sound.scp`
  - `data.sample_rate`: default `24000`
  - `data.segment_size`: default `240000`
- **Pretrained SSL (WavLM)**
  - `model.wavlmloss.ckpt_path`: default `pretrain_model/ssl/wavlm-base-plus`
  - `wav_lm_model`: default `pretrain_model/ssl/wavlm_model/wavlm`
- **Output**
  - `train.save_dir`: default `./exps/omnicodec`

## Training

Run training with the provided config:

```bash
python train.py -c config/config_omnicodec.yaml
```

Checkpoints and logs are written to `train.save_dir` (default: `./exps/omnicodec`).

## Inference (reconstruction)

### Prepare checkpoint

`infer.py` loads the checkpoint from:

- `pretrained_model/omnicodec.pth`

Place your pretrained weights at that path (or edit `infer.py` to point to your checkpoint).

### Run

Put test audio files in:

- `./testset/speech/`

Then run:

```bash
python infer.py -c config/config_omnicodec.yaml
```

Outputs will be written to:

- `./outputs/`

## Project structure

```text
.
├─ config/
│  └─ config_omnicodec.yaml
├─ dataset.py
├─ train.py
├─ infer.py
├─ models/
├─ modules/
├─ quantization/
├─ discriminators/
├─ losses/
├─ utils/
└─ requirements.txt
```

## Acknowledgements

- This repo benefits from [moshi](https://github.com/kyutai-labs/moshi)
- This repo benefits from [Qwen3Omni](https://github.com/QwenLM/Qwen3-Omni)
- This repo benefits from [DAC](https://github.com/descriptinc/descript-audio-codec)
- This repo benefits from [BigVGAN](https://github.com/NVIDIA/BigVGAN)
- This repo benefits from [SpeechTokenizer](https://github.com/zhangxinfd/speechtokenizer)

## Citation

If you use this work, please cite:

```bibtex
@misc{hu2026omnicodeclowframerate,
      title={OmniCodec: Low Frame Rate Universal Audio Codec with Semantic-Acoustic Disentanglement}, 
      author={Jingbin Hu and Haoyu Zhang and Dake Guo and Qirui Zhan and Wenhao Li and Huakang Chen and Guobin Ma and Hanke Xie and Chengyou Wang and Pengyuan Xie and Chuan Xie and Qiang Zhang and Lei Xie},
      year={2026},
      eprint={2603.20638},
      archivePrefix={arXiv},
      primaryClass={eess.AS},
      url={https://arxiv.org/abs/2603.20638}, 
}
```

## License

See the repository license