AFC-SPEX

Source code for the AFC-SPEX project.
Dataset and test samples are available at: https://box.nju.edu.cn/d/cd07cf71914c4edfa128/ (Password: afcspex1014)

About The Project

This repository contains the implementation of the AFC-SPEX algorithm, which is designed for combining multichannel adaptive feedback cancellation with speaker extraction.
Codes are being rebuilt: the baseline ideal_afc_dnsf and the proposed afc_spex are both available. More baselines will be added soon; they should be straightforward to implement within the current framework.
These codes are developed by H.C. Guo and Z. Li.

Use the environment.yml file to create the conda environment.
Download the dataset from https://box.nju.edu.cn/d/cd07cf71914c4edfa128/ (Password: afcspex1014), dataset name is prep.zip.
train.py is the main script for training the model (distributed training supported).
infer.py is the main script for inference; a closed-loop simulation is implemented there.
Run evaluate.py to evaluate the results.
Configuration files are in the configs/ folder.