search

BUTSpeechFIT / VBx

Variational Bayes HMM over x-vectors diarization
252 stars 57 forks source link

Spk loop elimination #33

Closed videodanchik closed 3 years ago

videodanchik commented 3 years ago

This PR includes speaker inner loop elimination (excluding the last line where I got a small time performance degradation in case of vectorizing it too). Also added some code style refactoring, following PEP recommendations.

Jamiroquai88 commented 3 years ago

LGTM

videodanchik commented 3 years ago

Added fastcluster and fixed small bug with missing parentheses.

Jamiroquai88 commented 2 years ago

@videodanchik @fnlandini Hey guys, I was running some experiments with the modified clustering (using fastcluster) and for larger dataset I got quite a different results compared to AHC. Did you have a chance to try this on a larger dataset? My dummy test was fine, so not sure what is going on, maybe I have a bug in my implementation (I integrated the change into my own codebase). Thanks.

fnlandini commented 2 years ago

Hi @Jamiroquai88 We verified that we obtain the same results for Callhome, DIHARD2 and both sets of AMI with VBx and we get the very same numbers. We did not check using only AHC, is there where you find differences?

Jamiroquai88 commented 2 years ago

Yeah, when running only AHC there are some differences in my tests. But nothing huge, ~1% difference in DER, so it is probably fine. Thanks anyway.

videodanchik commented 2 years ago

Hi @Jamiroquai88, could you please share the PLDA pairwise similarity matrix where you obtain different results while doing AHC with fastcluster and with your implementation? And, probably some examples of code where you output different speaker labels?

Jamiroquai88 commented 2 years ago

Sorry @videodanchik - it is gone. I did some other modifications after that and it looks like I had some other bug in the code, right now it looks fine. I mainly wanted to ask if @fnlandini tried it on some larger test sets, which he did (nice job).