Open PanYenChou opened 1 year ago
Our training dataset mainly consists of the wideband (16kHz) data from the DNS Challenge at InterSpeech 2021. The clean dataset used is "readbook" and the noise dataset incorporates the complete set of noises. Our test dataset utilizes the referenced test set released in the DNS Challenge at InterSpeech 2020.
Did you use drop band that reaches the performance in your paper?
The configs of my final reported results are listed in config. If there is a slight difference in your results, I think it's likely to be a randomness difference due to the dynamic mixing strategy. Besides, the type of GPU also has some effect on the results.
I'm trying to retrain your InterSubnet. I have some questions about training datasets.