utt-utt scoring using ivector-plda-scoring

62 views
Skip to first unread message

Divyesh Rajpura

unread,
Apr 18, 2020, 3:29:20 PM4/18/20
to kaldi-help
Currently, I am working on the Far-Field Speaker Verification Challenge. As there is no baseline is provided, I am using the aishell recipe as baseline. In challenge, we have been given trial files which are in form of utt-utt not spk-utt, so we have to consider given single enrolment utterance as only available data for modelling the speaker. But, ivector-plda-scoring considers all utts available for speaker and create speaker model. Is there any way I can perform utt-utt scoring?

Daniel Povey

unread,
Apr 19, 2020, 5:08:12 AM4/19/20
to kaldi-help
It would just be a case where you have one utterance per speaker, it's not fundamentally different from normal.


On Sun, Apr 19, 2020 at 3:29 AM Divyesh Rajpura <divyesh.r...@gmail.com> wrote:
Currently, I am working on the Far-Field Speaker Verification Challenge. As there is no baseline is provided, I am using the aishell recipe as baseline. In challenge, we have been given trial files which are in form of utt-utt not spk-utt, so we have to consider given single enrolment utterance as only available data for modelling the speaker. But, ivector-plda-scoring considers all utts available for speaker and create speaker model. Is there any way I can perform utt-utt scoring?

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/4dac2e39-1ef6-4b1b-89a7-9c6e64f32b58%40googlegroups.com.

Divyesh Rajpura

unread,
Apr 19, 2020, 6:24:44 AM4/19/20
to kaldi-help
@dan_povey, Thank you for your reply.

I have number of utterances for each speaker in enrolment. But, I have to consider only given single utterance as enrolment 

0373-F0373_373I0.25M_1_0192_normal 0538-T0538_538PCMR3M_recorded10_0135_normal
0373-F0373_373I0.25M_1_0192_normal 0521-S0521_521PCM-1.5M_recorded10_0167_normal
0351-T0351_351I0.25M_1_0092_normal 0540-F0540_540PCMR3M_recorded10_0133_normal
0351-T0351_351I0.25M_1_0092_normal 0547-F0547_547PCMR3M_recorded10_0147_normal
0355-S0355_355I0.25M_1_0170_normal 0398-T0398_398PCMR3M_recorded10_0199_normal
0355-S0355_355I0.25M_1_0170_normal 0396-F0396_396PCMR3M_recorded14_0150_normal

This is the portion of trial file. Each value in that represents utt-id. First part of each utt-id represents spk-id (i.e. 0373, 0538, etc.). First column is enrolment utt and second column is test utt. As we can see I have multiple utt per speaker but at a time I have to consider only single utt to modeling the speaker.

Currently, what ivector-plda-scoring is doing is it takes all utts of particular spk and calculate single representation of that speaker and use that. That is not the case in our evaluation strategy. 



On Sunday, April 19, 2020 at 2:38:12 PM UTC+5:30, Dan Povey wrote:
It would just be a case where you have one utterance per speaker, it's not fundamentally different from normal.


On Sun, Apr 19, 2020 at 3:29 AM Divyesh Rajpura <divyesh.r...@gmail.com> wrote:
Currently, I am working on the Far-Field Speaker Verification Challenge. As there is no baseline is provided, I am using the aishell recipe as baseline. In challenge, we have been given trial files which are in form of utt-utt not spk-utt, so we have to consider given single enrolment utterance as only available data for modelling the speaker. But, ivector-plda-scoring considers all utts available for speaker and create speaker model. Is there any way I can perform utt-utt scoring?

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi...@googlegroups.com.

Daniel Povey

unread,
Apr 19, 2020, 6:28:47 AM4/19/20
to kaldi-help
OK, well just give it the ones you need.

To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/2a563dc2-af0e-4574-a002-233ee69c835e%40googlegroups.com.

Divyesh Rajpura

unread,
Apr 19, 2020, 1:55:40 PM4/19/20
to kaldi-help
I have changed the input for enrolment in ivector-plda-scoring and now its working perfectly. Thank You.
Reply all
Reply to author
Forward
0 new messages