Full-band and Narrow-band fusion Network for SSL

Introduction

This repository provides methods which based on full-band and narrow-band fusion network for sound source localization. The narrow-band module processes the along-time sequences to focus on learning these narrow-band spatial information. The full-band module processes the along-frequency sequence to focus on learning the full-band correlation of spatial cues, such as the linear relation of DP-IPD to frequency.

Methods

Thress official implemented sound source localization methods are included:

• "FN-SSL: Full-Band and Narrow-Band Fusion for Sound Source Localization" , INTERSPEECH, 2023.
• "IPDnet: A Universal Direct-Path IPD Estimation Network for Sound Source Localization" TASLP, 2024
• "IPDnet2: an efficient and improved inter-channel phase difference estimation network for sound source localization"

Datasets

• Source signals: from LibriSpeech database
• Noise source signals: from Noise92
• Real-world multi-channel microphone signals: from LOCATA database
• RealMAN dataset: from RealMAN

Quick start (will be update soon)

• Preparation

  • Download the required dataset and organize the data according to the data_org in the data folder.
  • Generate multi-channel data, You can set data_num (in Simu.py) to control the size of the dataset. --train, -- test, --dev are used to control the generation of train dataset, test dataset, and validation dataset, respectively. The source data path of them are specified by dirs ['sousig_train '] in Opt.py.

  python Simu.py --train/--test/--dev
  

• Training

  • We have implemented both FN-SSL and IPDnet using the Pytorch-lightning framework.
  • For Train,

  python main.py fit --data.batch_size=[*,*] --trainer.devices=*,*
  

  • For test,

  python main.py test  --ckpt_path logs/MyModel/version_x/checkpoints/**.ckpt --trainer.devices=*,*
  

• Pretrained models

  • Using the FN_lightning model to load the lightning checkpoint in torch framework.

Framework	Task	Checkpoint
Lightning	DP-IPD regression (FN-SSL)	https://pan.baidu.com/s/1zRKpiqbSuo80Xu5ZRoS1gQ?pwd=6w51
Lightning	DOA classification (FN-SSL)	https://pan.baidu.com/s/1U1Wl5ZBZBItc2Vku7AyqNA?pwd=ceqm

more checkpoints will be update soon.

Citation

If you find our work useful in your research, please consider citing:

@InProceedings{wang2023fnssl,
    author = "Yabo Wang and Bing Yang and Xiaofei Li",
    title = "FN-SSL: Full-Band and Narrow-Band Fusion for Sound Source Localization",
    booktitle = "Proceedings of INTERSPEECH",
    year = "2023",
    pages = ""}

Reference code

• SRP-DNN
• Cross3D
• icoCNN

We gratefully acknowledge the Cross3D and icoCNN, which provided the foundation for the dataset generation and simulation procedures used here.

Licence

MIT