Model Explainability

[PaulSZH95]

May I check with you, for the vad tutorial provided in the repo, 2s chunks were used for training. Could I check if the resolution is 16ms?

I am trying to visualize the effect of the sincnet layer and am attempting to extract the sincnet ParamSincFB layer. However, the VAD pipeline for inference breaks for this purpose and I am trying to replicate the Pipeline method with torchaudio.load instead. Could I check if I should use 16ms or 2 s.

[PaulSZH95]

Hi @hbredin much apologies for the tag as the issue seems more related to product knowledge than technical know-hows. In the supplementary information, the model summary is as follows and suggest that the model ingests 2s chunks directly. If so, could I check at what stage does the mili-second segmentation starts?:

Truly much thanks.