Prem

Hi, Library version that we used was 20.1 for Catalyst

Hi, I did not solve the issue. I tried with around 20k audio clips for 2 person speech separation only. I would assume now that more data than that would...

You mean to say while preparing the dataset? Well, I've seen someone mention that [here](https://www.reddit.com/r/MachineLearning/comments/dpsy5l/d_i_have_acquired_the_videos_to_the_avspeech/). However, adding additional noise like AudioSet might help regularise.

Yes, you could make out in certain instances who was the main speaker in the separated output. But, not always. Sometimes, it was only noise or mix of both the...

Hi @MordehayM , I believe it was 20k unique clips. 200C2 is indeed 19k, however not all combinations are considered. There is a parameter: REMOVE_RANDOM_CHANCE (in audio_mixer_generator.py). This will prevent...

Hi, Thanks for pointing out the mistake. It was fixed when the repo was merged to asteroid [here](https://github.com/asteroid-team/asteroid/blob/master/egs/avspeech/looking-to-listen/model.py#L297), somehow it was not fixed here. But I did not see any...

I see... we are mixing the axis at the last layer. Thank you again for pointing this out! I will fix this.

[Here](https://github.com/vitrioil/Speech-Separation/blob/master/src/models/models.py#L231) input channel expected by `self.conv1` is 512 so transposing the channels so that we get the correct expected channel, otherwise channels will still be 1. I don't have access...

Hi @MordehayM, I still haven't tried re-training, so currently the result is the same