search

Audio-WestlakeU / NBSS

The official repo of NBC & SpatialNet for multichannel speech separation, denoising, and dereverberation
MIT License
175 stars 21 forks source link

dataset #10

Open PriyankaPaud opened 1 year ago

PriyankaPaud commented 1 year ago

Where can I find the WSJ0 dataset to run the project? Can I use other datasets instead of WSJO ?

PriyankaPaud commented 1 year ago

How can I convert .wv1 files into .wav file?

quancs commented 1 year ago

Where can I find the WSJ0 dataset to run the project? Can I use other datasets instead of WSJO ?

You can download the WSJ0 dataset from https://catalog.ldc.upenn.edu/LDC93S6A

quancs commented 1 year ago

How can I convert .wv1 files into .wav file?

You can use the sph2pipe tool sph2pipe.zip

PriyankaPaud commented 1 year ago

I can't find the model in the github repo . I want to test the model . Seems like the pushing of model is prohibited by .gitignore file. Can anyone help me with the model file ?

quancs commented 1 year ago

What do you mean the model?

PriyankaPaud commented 1 year ago

I mean the actual model the trained one by the team ?

quancs commented 1 year ago

The trained models are not uploaded in the repo.

PriyankaPaud commented 1 year ago

Can I get access to the trained model ? It would be a great help

quancs commented 1 year ago

Which model you want?

quancs commented 1 year ago

Accutally, you can train the models, like NB-BLSTM, NBC2, using the scripts provided in the repo from scratch.

PriyankaPaud commented 1 year ago

Yeah will be doing that once we test on our data . For testing purposes we need the model at first

quancs commented 1 year ago

The model we trained were using our private code (we cleaned the code by removing, renaming some classes or functions), so it might take you some additional work to load the weights. Or you can train the model from scratch, so that the obtained model will match the code.

NBC2_small - epoch99_neg_si_sdr-19.0141.zip

PriyankaPaud commented 1 year ago

Okay thankyou so much for this .

quancs commented 1 year ago

You're welcome.

PriyankaPaud commented 1 year ago

Do we need to include al dataset provided by WSJ0 while training? or only few will also work ?

quancs commented 1 year ago

Not all the data in WSJ0, only the utterances used in WSJ0-2mix

PriyankaPaud commented 1 year ago

Can you share the dataset the unmixed ones which can be used easily to train the script. I see that the dataset path and the dataset name all are hardcoded ? Can you help here . I just want the dataset that will help run the script .Thankyou . I have a dataset but the train process gets killed after 1 epoch only.

PriyankaPaud commented 1 year ago

Do you have the list of utterances which were used during training ? Its very hard to analyze from the wsj0-mix