HybridBeam: Hybrid Neural Networks for On-Device Directional Hearing

This is the source code for our AAAI 22 paper: Hybrid Neural Networks for On-Device Directional Hearing.

To generate the synthetic datasets, you need to first download the original speech and noise datasets by running download_datasets.sh. The network is defined in deepbeam.py. To generate datasets, run train.py data; then to train the model, run train.py train; lastly, to evaluate, run train.py test. We also attached a pretrained model for 6-mic DeepBeam+ model in pretrained pretrained_6mic.bin.

The code requires a C/C++ implementation of WebRTC beamformer. The code is in the beamformer folder where you can compile following the README.md there. and a compiled x86-64 version is beamform_mic_array where you can directly call using the Python code.