search

fakufaku / torchiva

Blind source separation with independent vector analysis family of algorithm in torch
https://torchiva.readthedocs.io/en/latest/
MIT License
86 stars 5 forks source link

Trained model #6

Closed aleksandra-bebe closed 9 months ago

aleksandra-bebe commented 10 months ago

Hello, does this trained model work just with tiss or can work with mvdr and others beamformers?

fakufaku commented 10 months ago

Hi, I haven't tried, but I think it would work just with tiss.

aleksandra-bebe commented 10 months ago

Thank you for answer, yes it only works with tiss, and just one more question is script test_beamformers use model or is that just to see if the beamformers work and if we want to train models for beamformers should the dataset you specified for training need to be paid for?

fakufaku commented 9 months ago

test_beamformers.py is just a coffee correctness test and does not use a trained model. But you could use it as a reference on how to use the code.

The dataset we used was based on WSJ0, which is paid. The noise part is from Chime3 challenge and I believe it's free. You could replace the speech by that of librimix for example if you need something free. It's structure is similar to WSJ0, so should require minimal code change.

aleksandra-bebe commented 9 months ago

Thank you very much.