Getting Error when using pre trained embedddings to extract ivector #4496
Comments
|
Likely either a feature dimension mismatch (13 vs. 40), or a splicing
mismatch, e.g. the number of frames to splice is wrongly specified. If the
latter, it would likely mean you had
generated the config files for ivector stuff with the wrong inputs.
…On Wed, Apr 14, 2021 at 4:57 PM Tushar Palawat ***@***.***> wrote:
I was trying to extract ivector from pre-trained librispeech model and
running the script to extract ivectors, I am getting this error. Does
anyone have any idea as to why it is happening? Just to mention, I am also
getting the same error in my log file with voxceleb model too, if this
information helps.
Please let me know in case I need to post additional information, Thanks!
copy-feats --compress=true ark:- ark,scp:24_ivector_extraction/wsj_SI284_dot_cor.v2/ivector_wsj_SI284_dot_cor.v2_online.5.ark,24_ivector_extraction/wsj_SI284_dot_cor.v2/ivector_wsj_SI284_dot_cor.v2_online.5.scp
ivector-extract-online2 --config=0013_librispeech_v1_extractor/ivector_extractor/../iv_conf/ivector_extractor.conf ark:05_feat/wsj_SI284_dot_cor.v2/split20/5/spk2utt scp:/share/mini1/res/t/asr/studio/read-us/wsj/embeddings/exp/exp_embedding/05_feat/wsj_SI284_dot_cor.v2/split20/5/feats.scp ark:-
LOG (ivector-extract-online2[5.5.1048~1-d211d]:ComputeDerivedVars():ivector-extractor.cc:183) Computing derived variables for iVector extractor
LOG (ivector-extract-online2[5.5.1048~1-d211d]:ComputeDerivedVars():ivector-extractor.cc:204) Done.
ASSERTION_FAILED (ivector-extract-online2[5.5.1048~1-d211d]:Check():online-ivector-feature.cc:87) Assertion failed: (lda_mat.NumCols() == spliced_input_dim || lda_mat.NumCols() == spliced_input_dim + 1)
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
<#4496>, or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AAZFLOZKALRW36MJM2TGNL3TIVKHRANCNFSM4246HQOA>
.
|
|
Hi, so I made some changes in my script. Earlier, I wasn't using splice-feats.cc binary and then transforming the matrix. And then further in the log file, I get the final error. I have 30 dimensions in the MFCC features I created vs the features in librispeech model. Could it be because of that? |
|
You shouldn't be using splice-feats and should not have to change the
script.
This like likely about regular (13-dim) vs. "hires" (40-dim) MFCCs.
…On Thu, Apr 15, 2021 at 7:17 PM Tushar Palawat ***@***.***> wrote:
Hi, so I made some changes in my script. Earlier, I wasn't using
splice-feats.cc binary and then transforming the matrix.
I have now done that, but I am now getting the following warnings.
WARNING (transform-feats[5.5.1048~1-d211d]:main():transform-feats.cc:110) Transform matrix for utterance WSJ-F46n-46nc020u_1:0-15744 has bad dimension 40x281 versus feat dim 91
LOG (transform-feats[5.5.1048~1-d211d]:main():transform-feats.cc:161) Applied transform to 0 utterances; 1666 had errors.
WARNING (transform-feats[5.5.1048~1-d211d]:main():transform-feats.cc:110) Transform matrix for utterance WSJ-F46n-46nc020v_1:0-12051 has bad dimension 40x281 versus feat dim 91
WARNING (ivector-extract[5.5.1048~1-d211d]:RunPerSpeaker():ivector-extract.cc:123) No features present for utterance WSJ-F46c-46cc0201_1:0-10267
And then further in the log file, I get the final error.
LOG (apply-cmvn-online[5.5.1048~1-d211d]:main():apply-cmvn-online.cc:133) Applied online CMVN to 1666 files, or 1289059 frames.
WARNING (transform-feats[5.5.1048~1-d211d]:main():transform-feats.cc:110) Transform matrix for utterance WSJ-F46p-46pc041f_1:0-8220 has bad dimension 40x281 versus feat dim 91
LOG (transform-feats[5.5.1048~1-d211d]:main():transform-feats.cc:161) Applied transform to 0 utterances; 1666 had errors.
LOG (gmm-global-get-post[5.5.1048~1-d211d]:main():gmm-global-get-post.cc:115) Done 0 files, 0 with errors, average UBM log-likelihood is -nan over 0 frames.
WARNING (gmm-global-get-post[5.5.1048~1-d211d]:Close():kaldi-io.cc:515) Pipe apply-cmvn-online --spk2utt=ark:05_feat/wsj_SI284_dot_cor.v2/split20/4/spk2utt --config=0013_librispeech_v1_extractor/ivector_extractor/online_cmvn.conf 0013_librispeech_v1_extractor/ivector_extractor/global_cmvn.stats scp:05_feat/wsj_SI284_dot_cor.v2/split20/4/feats.scp ark:- | splice-feats --left-context=3 --right-context=3 ark:- ark:- | transform-feats 0013_librispeech_v1_extractor/ivector_extractor/final.mat ark:- ark:- | had nonzero return status 256
ERROR (gmm-global-get-post[5.5.1048~1-d211d]:~SequentialTableReaderArchiveImpl():util/kaldi-table-inl.h:678) TableReader: error detected closing archive 'apply-cmvn-online --spk2utt=ark:05_feat/wsj_SI284_dot_cor.v2/split20/4/spk2utt --config=0013_librispeech_v1_extractor/ivector_extractor/online_cmvn.conf 0013_librispeech_v1_extractor/ivector_extractor/global_cmvn.stats scp:05_feat/wsj_SI284_dot_cor.v2/split20/4/feats.scp ark:- | splice-feats --left-context=3 --right-context=3 ark:- ark:- | transform-feats 0013_librispeech_v1_extractor/ivector_extractor/final.mat ark:- ark:- |'
I have 30 dimensions in the MFCC features I created vs the features in
librispeech model. Could it be because of that?
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#4496 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AAZFLO3ZNJOZHMKXAR3MA3DTI3DN7ANCNFSM4246HQOA>
.
|
|
@danpovey Thank you very much. It was indeed a problem of dimension mismatch. There were a few more errors in the script but it did work in the end. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
I was trying to extract ivector from pre-trained librispeech model and running the script to extract ivectors, I am getting this error. Does anyone have any idea as to why it is happening? Just to mention, I am also getting the same error in my log file with voxceleb model too, if this information helps.
Please let me know in case I need to post additional information, Thanks!
The text was updated successfully, but these errors were encountered: