Neural Network based Sound Source Localization Models

This repository includes the programs to run and test the neural network models that we proposed in our publications on sound source localization, including

• Deep learning based direction-of-arrival (DOA) estimation
• Joint DOA estimation and speech/non-speech classification

The networks has been trained to work with the microphone array of the Softbank Pepper robot (early version with directional microphones).

Please cite the relevant publications when using the code.

Dependency

• Python <https://www.python.org/>_ (version >= 3.8)
• NumPy <http://www.numpy.org/>_ (version >= 1.14)
• PyTorch <https://pytorch.org/>_ (version >= 0.2)
• apkit <https://github.com/hwp/apkit>_ (version >= 0.2)

Data

We use the SSLR dataset <https://www.idiap.ch/dataset/sslr>_ for the experiments.

Usage

Training

---

This command reproduces the "ResNet-STFT" model described in "Deep Learning Approaches for Auditory Perception in Robotics", Weipeng He, PhD Thesis, EPFL <https://infoscience.epfl.ch/record/283940>_:

.. code::

ssl_nn_v2/train_resnet_2stage.py --model=models/test --window-size=8192 --hop-size=4096 --n-doa=360 --output-act=5 --n-out-hidden=1 --sigma=8 --s1-epoch=4 --epoch=10 --lr=0.001 --ld=2 --batch-size=128 <SSLR_PATH>/lsp_train_*

Inference

---

.. code::

ssl_nn_v2/test_nn_raw.py --feature=stft --method=METHOD_NAME --window-size=8192 --hop-size=4096 --batch-size=100 <SSLR_PATH>/human models/thesis_resnet_act5_p1lsp_s1ep4_ep10_valid_b100

Evaluation

---

.. code::

eval/gen_2tasks_report.py --method=METHOD_NAME --window-size=8192 --hop-size=4096 --output=REPORT_DIR --ssl-only <SSLR_PATH>/human gnuplot --persist REPORT_DIR/ssl_pr_plot

Publications

The models and code in this repository are based on the work published in:

Deep Neural Networks for Multiple Speaker Detection and Localization Weipeng He, Petr Motlicek, Jean-Marc Odobez In IEEE International Conference on Robotics and Automation (ICRA), 2018

Joint Localization and Classification of Multiple Sound Sources Using a Multi-task Neural Network Weipeng He, Petr Motlicek, Jean-Marc Odobez In INTERSPEECH, 2018

Deep Learning Approaches for Auditory Perception in Robotics Weipeng He PhD Thesis, EPFL