SRP-DNN

A python implementation of “SRP-DNN: Learning direct-path phase difference for multiple moving sound source localization”, IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022.

• Contributions
  • Learning competing and time-varying direct-path inter-channel phase differences (or IPD sequence) for multiple moving sources
    • avoids the assignment ambiguity and the problem of uncertain output-dimension encountered when simultaneously predicting multiple targets
    • exhibits reliable peaks around the actual directions of sources by the constructed spatial spectrum
    
  • Iterative source detection and localization
    • separates the merged peaks of spatial spectrum caused by the interaction between sources
    • achieves superior performance for the azimuth and elevation estimation of multiple moving sound sources
    
• Suited cases
  • good or adverse noisy and reverberant scenario
  • single or multiple sound sources
  • static or moving source sources
  • the number of sound sources is known or unknown
  • different topologies of microphone arrays

Datasets

• Source signals: from LibriSpeech database
• Real-world multi-channel microphone signals: from LOCATA database

Quick start

• Preparation

  • copy the train-clean-100, dev-clean and test-clean folders of LibriSpeech database to SRP-DNN/data/SrcSig/LibriSpeech
  • install: numpy, scipy, soundfile, tqdm, matplotlib, gpuRIR, webrtcvad, etc.

• Training

  python RunSRPDNN.py --train --gen-on-the-fly --gpu-id [*] (--use-amp)
  

• Evaluation

  • use GPU

  python RunSRPDNN.py --test --gpu-id [*] --time 00000001 --eval-mode locata pred eval (--use-amp)
  

  • use CPU

  python RunSRPDNN.py --test --no-cuda --time 00000001 --eval-mode locata pred eval (--use-amp)
  

• Pretrained models

  • exp/00000002/best_model.tar

Citation

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

@InProceedings{yang2022srpdnn,
    author = "Bing Yang and Hong Liu and Xiaofei Li",
    title = "SRP-DNN: Learning direct-path phase difference for multiple moving sound source localization",
    booktitle = "Proceedings of {IEEE} International Conference on Acoustics, Speech and Signal Processing (ICASSP)",
    year = "2022",
    pages = "721-725"}

Reference code

• Cross3D

Licence

MIT