Checkpoints and RIRs

Checkpoints

The release page has two model checkpoints. All checkpoints include “model_state_dict”, “optimizer_state_dict”, and some other meta information.

The first model checkpoint is the original model checkpoint at the 58th epoch. The performance is shown in this table:

	With Reverb				No Reverb
Method	WB-PESQ	NB-PESQ	SI-SDR	STOI	WB-PESQ	NB-PESQ	SI-SDR	STOI
FullSubNet	2.987	3.496	15.756	0.926	2.889	3.385	17.635	0.964

In addition, some people are interested in the performance when using cumulative normalization. The below one is a pre-trained FullSubNet using cumulative normalization:

	With Reverb				No Reverb
Method	WB-PESQ	NB-PESQ	SI-SDR	STOI	WB-PESQ	NB-PESQ	SI-SDR	STOI
FullSubNet (Cumulative Norm)	2.978	3.503	15.820	0.928	2.863	3.376	17.913	0.964

If you want to inference or fine-tune based on these checkpoints, please check the usage in the documents.

Room Impulse Responses

As mentioned in the paper, the room impulse responses (RIRs) come from the Multichannel Impulse Response Database and the Reverb Challenge dataset. Please download the zip package “RIR (Multichannel Impulse Response Database + The REVERB challenge).zip” from the release page if you would like to retrain the FullSubNet.

Note that the zip package includes a folder “rir” and a file “rir.txt.” The folder “rir” contains all separated single-channel RIRs extracted from the above two datasets. The suffix (e.g., “m_”) of the filename is the index of a microphone. The file “rir.txt” is just a path list of all RIRs. Please modify it to fit your case before you use it.

For some cases, if you would like to extract channel by yourself, you can download these RIRs from pages:

1. Multichannel Impulse Response Database

2. The REVERB challenge data

   1. https://reverb2014.dereverberation.com/tools/reverb_tools_for_Generate_mcTrainData.tgz

   2. https://reverb2014.dereverberation.com/tools/reverb_tools_for_Generate_SimData.tgz