Integrate SV Model about similaritylearning HOT 3 CLOSED

The best performance I got so far on VoxCeleb is with ClopiNet architecture on top of MFCC features.
A tutorial and pre-trained model is available here

The issue with this model is that it relies on MFCC which are slow to compute: one cannot really precompute them because random noise is added on-the-fly to the audio file as a data augmentation step. Therefore, I have been trying recently to replace the MFCC features by trained SincNet features (computed from the waveform directly) but to no avail so far.

Yet, I'd rather have you work directly from the waveform (looks cleaner).
So, unless I manage to switch ClopiNet to waveform quickly, I suggest you use the architecture provided in the SincNet repo.

from similaritylearning.

@hbredin Got it. A priori I'll work on integrating SincNet

from similaritylearning.

SincNet is working using our custom loss modules, like arc margin or congenerous cosine (CoCo).
A better implementation will probably come once VoxCeleb is integrated.

from similaritylearning.