News

Now you can train multiple models including sepformer and mossformer2 using this dataset with speechbrain here. You can also checkout our paper for more details.

Aishell1Mix

Aishell1Mix is a mandarin version of speech separation dataset like WSJMix and LibriMix. It mixes 2 or 3 speaker sources from the open source mandarin speech corpus Aishell1 with the noise dataset WHAM. The scripts are modified from LibriMix. Please refer to it for more details.

How to generate

Firstly make sure that SoX is installed on your machine.

• For windows :

conda install -c groakat sox

• For Linux or MacOS:

conda install -c conda-forge sox

Then to generate LibriMix, clone the repo and run the main script: generate_aishell1mix.sh

git clone https://github.com/huangzj421/Aishell1Mix.git
cd Aishell1Mix
pip install -r requirements.txt
./generate_aishell1mix.sh storage_dir

Features

In Aishell1Mix you can choose :

• The number of sources in the mixtures.
• The sample rate of the dataset from 16 KHz to any frequency below.
• The mode of mixtures : min (the mixture ends when the shortest source ends) or max (the mixtures ends with the longest source)
• The type of mixture : mix_clean (utterances only) mix_both (utterances + noise) mix_single (1 utterance + noise)

You can customize the generation by editing generate_aishell1mix.sh.

Citing

Please, cite Aishell1Mix if you use it for your research or business.

@inproceedings{huang2025aishell1mix,
  title={Aishell1Mix: Towards Robust Mandarin Speech Separation with Scalable Audio Language Models},
  author={Huang, Zijian and Subakan, Cem},
  booktitle={National Conference on Man-Machine Speech Communication},
  pages={187--200},
  year={2025},
  organization={Springer}
}