Geometry-Agnostic Acoustic Processing: A Dynamic Spatial Network for Joint Echo Cancellation and Noise Suppression

Requirements

pip install -r requirements.txt

Train & Test

This project is built on the pytorch-lightning package.

Train

python Trainer.py fit \
 --config=configs/config.yaml \ # network config
 --model.arch.dim_output=2 \ # output dim per T-F point
 --model.arch.num_freqs=129 \ # the number of frequencies, related to model.stft.n_fft
 --data.train_dir=/datasets/train \ # the path of train dataset
 --data.test_dir=/datasets/val \ # the path of val dataset
 --data.batch_size=[8,16] \ # batch size for train and val
 --trainer.devices=0, \ # train device
 --trainer.max_epochs=100 # better performance may be obtained if more epochs are given

Test the model trained:

python Trainer.py test --config=logs/VSAECNet/version_x/config.yaml \ 
 --ckpt_path=logs/VSAECNet/version_x/checkpoints/epochY.ckpt \ 
 --trainer.devices=0,

Test Result Demonstration (Demo)

The code repository includes an interactive demo with test audio samples in the wav directory, demonstrating the system's core capabilities. Using a single trained checkpoint, we evaluate performance across three key scenarios: far-end single-talk (FST), near-end single-talk (NST), and double-talk (DT). For the DT scenario specifically, we provide acoustic echo cancellation (AEC) and noise suppression results for three different microphone array configurations (3-, 4-, and 6-microphone setups).