Robust Sound Source Tracking Using SRP-PHAT and 3D Convolutional Neural Networks

Code repository for the paper Robust Sound Source Tracking Using SRP-PHAT and 3D Convolutional Neural Networks [1].

• Dependencies
• Datasets
• Other source files
• Pretrained models
• References

Dependencies

• Python: it has been tested with Python 3.8
• Numpy, matplotlib, scipy, soundfile, pandas and tqdm
• Pytorch: it has been tested with Python 1.4.0
• webrtcvad
• gpuRIR [2]

Datasets

• LibriSpeech The training dataset is generated during the training of the models as the trajectories are needed, but to simulate them you will need to have the LibriSpeech corpus in your machine. By default, the main scripts look for it in datasets/LibriSpeech but you can modify its phat with the path_train and path_test variables.
• LOCATA In order to test the models with actual recordings, you will also need the dataset of the LOCATA challenge. By default, the main scripts look for it in datasets/LOCATA but you can modify its phat with the path_locata variable.

Main scripts

There is a main script for each model analyzed in the paper. The scripts train the model, save the trained model, test it for several reverberation times and SNRs, save the results, plot them, and finally test the model with the recordings of the LOCATA challenge.

• 1sourceTracking_Cross3D.py Cross3D is the main contribution of the paper, performing causal DOA estimation and tracking of one acoustic source using a 3D CNN over sequences of SRP-PHAT power maps. You can modify the resolution of the power maps with the res_the and res_phi variables.
• 1sourceTracking_fromGCCs.py This model uses a causal 1D CNN over GCCs sequences. Although using GCCs as inputs of a Neural Network for DOA estimation is a quite common technique, using 1D CNNs over them to also perform the tracking of the source is an approach, to the best of the authors' knowledge, unpublished.
• 1sourceTracking_fromMaximums.py This model uses a causal 1D CNN over the coordinates of the maximums of high resolution SRP-PHAT maps.
• 1sourceTracking_fromSpectrograms.py This model uses a causal 2D CNN over the spectrograms of the signals of each microphone.
• 1sourceTracking_SELDnet.py This model is a modification of the proposed in [3], we removed the Sound Event Detection (SED) output and keep only one localization output since we were interested only in the tracking of one source. It uses a combination of 2D convolutional layers and non-causal recurrent layers over the spectrograms of the signals of each microphone. You can find the whole model in its repository.

The RAM consumption keeps increasing during the training, but it should still be possible to train the models in most of the machines, although it may crash when it try to test them (after saving the trained model). Please, let me know if you find where the memory leak is.

The training of the models took about 25 hours in a Nvidia GeForce GTX 980Ti with an Intel Core i7-6700 CPU and 16GB of RAM. The test in the simulated scenarios took about 1 hour.

Pretrained models

You can find the pretrained models in the release section of the repository.

Other source files

acousticTrackingDataset.py, acousticTrackingLearners.py, acousticTrackingModels.py and acousticTrackingDataset.py contain several classes and functions employed by the main scripts. They have been published to facilitate the replicability of the research presented in [1], not as a software library. Therefore, any feature included in them that is not used by the main scripts may be untested and could contain bugs.

References

[1] D. Diaz-Guerra, A. Miguel, J.R. Beltran, "Robust Sound Source Tracking Using SRP-PHAT and 3D Convolutional Neural Networks," [arXiv preprint].

[2] D. Diaz-Guerra, A. Miguel, J.R. Beltran, "gpuRIR: A python library for Room Impulse Response simulation with GPU acceleration," [arXiv preprint].

[3] Sharath Adavanne, Archontis Politis, Joonas Nikunen, and Tuomas Virtanen, "Sound event localization and detection of overlapping sources using convolutional recurrent neural network" in IEEE Journal of Selected Topics in Signal Processing (JSTSP 2018).