This repo provides a reference implementation of DSPM as described in the paper:
Improving Monaural Speech Enhancement with Dynamic Scene Perception Module
Accpted by ICME, 2022.
105 types of noise are concatenated for training and validation, while 110 types of noise for testing (add 5 unseen noises). The ratio of the invisible part to the visible part of our test noise is about 4:1 (5 unseen noises vs. 105 seen noises, 20 minutes vs. 5 minutes)
The experimental platform is Ubuntu LTS 18.04 with i7-9700 and RTX 2060.
For any questions please open an issue or drop an email to: wxtai@std.uestc.edu.cn