danpovey
commented
3 years ago That looks like a combination of hires (40-dim) vs. normal (13-diim) MFCCs plus other configuration differences. If it was just for some utterances, you likely combined directories that had different MFCC dims.
On Fri, Apr 23, 2021 at 10:49 PM Tushar Palawat @.***> wrote:
I was trying to extract pretrained embeddings for voxceleb and I am getting an error which could probably signify feature mismatch. This is despite the fact I am using the same MFCC config used as specified in voxceleb MFCC config.
--sample-frequency=16000 --frame-length=25 # the default is 25 --low-freq=20 # the default. --high-freq=7600 # the default is zero meaning use the Nyquist (8k in this case). --num-mel-bins=30 --num-ceps=24 --snip-edges=false
The error that I am getting when extracting ivector is,
WARNING (transform-feats[5.5.1048~1-d211d]:main():transform-feats.cc:110) Transform matrix for utterance WSJ-F46a-46ac041d_1:0-10268 has bad dimension 200x401 versus feat dim 168 LOG (transform-feats[5.5.1048~1-d211d]:main():transform-feats.cc:161) Applied transform to 0 utterances; 1621 had errors. WARNING (ivector-extract[5.5.1048~1-d211d]:RunPerSpeaker():ivector-extract.cc:123) No features present for utterance WSJ-F40h-40hc0201_1:0-1895
I am getting this warning for multiple utterances, this is just a sample. LOG (ivector-extract[5.5.1048~1-d211d]:RunPerSpeaker():ivector-extract.cc:190) Done 0 speakers; 12 with errors. 0 utterances were processed, 1621 with errors. This is the final log I get for ivector-extract binary.
Also, If this bit of information helps, I have tried to use both splice-feats (with left and right context as 3) and add-deltas (using standard voxceleb delta config) but neither has worked and both threw errors. Please let me know if you need any further information, Thanks!
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/kaldi-asr/kaldi/issues/4505, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAZFLOZGZNPVGN2T7OGJ3HTTKGCHHANCNFSM43OZHBFA .
TPalawaT
kkm000
I was trying to extract pretrained embeddings for voxceleb and I am getting an error which could probably signify feature mismatch. This is despite the fact I am using the same MFCC config used as specified in voxceleb MFCC config.
The error that I am getting when extracting ivector is,
I am getting this warning for multiple utterances, this is just a sample.
LOG (ivector-extract[5.5.1048~1-d211d]:RunPerSpeaker():ivector-extract.cc:190) Done 0 speakers; 12 with errors. 0 utterances were processed, 1621 with errors.This is the final log I get for ivector-extract binary.Also, If this bit of information helps, I have tried to use both splice-feats (with left and right context as 3) and add-deltas (using standard voxceleb delta config) but neither has worked and both threw errors. Please let me know if you need any further information, Thanks!