CleanMel

Pytorch implementation of "CleanMel: Mel-Spectrogram Enhancement for Improving Both Speech Quality and ASR".

Paper 🤩 | Demos 🎶 | Issues 😅 | Lab 🙉 | Contact 😘

Noticement

• A quick inference demo is provided in shell/inference.sh.
• For the ASR inference codes, we released in the asr_infer branch.
• This repo is under development and will be formally released ASAP.

Introduction

Getting Started

Step 1. Pretrained checkpoint:

Download the pretrained checkpoints here (Google drive).

Step 2. Environment configurations:

Create a new conda environment for CleanMel:

conda create -n CleanMel python==3.10.14

Setup the required packages:

pip install -r requirements.txt

Hint: If you have any problem w.r.t. the environments. Please first make sure your mamba_ssm, torch and pytorch-lightning versions are identical to those in the requirements.txt.

Inference

Change pretrained checkpoint path:

In ./shell/inference.sh:

1. change the model.arch_ckpt to YOUR_PATH/CleanMel_S_L1.ckpt
2. change the model.vocos_ckpt to YOUR_PATH/vocos_offline.pt.

---

Inference provided demos:

Using the inference.sh in shell folder:

cd shell
bash inference 0,

Hint: the number behind indicates your GPU ID. 0, means using only GPU:0 for inference. And change it value to 0,1,... to allow more GPUS for inference.

---

Inference your own file:

In inference.sh:

1. change dataset.speech_dir to your data path.
2. change model.output_path to your output folder.

Run

bash inference 0,

Hint: You could also delete model.output_path, by default, the waveform will output to the same folder of your pretrained checkpoint.

Training

Performance

Speech enhancement performance

ASR performance

Citation

@misc{shao2025cleanmel,
    title={CleanMel: Mel-Spectrogram Enhancement for Improving Both Speech Quality and ASR}, 
    author={Nian Shao and Rui Zhou and Pengyu Wang and Xian Li and Ying Fang and Yujie Yang and Xiaofei Li},
    year={2025},
    eprint={2502.20040},
    archivePrefix={arXiv},
    primaryClass={eess.AS},
    url={https://arxiv.org/abs/2502.20040}
}