0:00:19PLDA based speaker verification with weighted LDA do techniques.
0:00:28This is the outline of my presentation. First part is motivation, where I will discuss
0:00:30why we have investigated different techniques such as LDA weighted with PLDA system.
0:00:43Using dimensionality reduction on i-vector features.
0:00:48I will then discuss the experiments on telephone and microphone speech with PLDA system
0:00:54which is based on LDA and weighted LDA dimension reduction techniques.
0:01:03Our main motivation in this paper is to identify the best channel compensation approach for
0:01:06telephone and microphone based speaker verification system.
0:01:12Dehak has investigated dimensionality reduction techniques for channel compensation is the i-vector system.
0:01:19And he has investigated PLDA modeling with i-vectors to compensate channel variability.
0:01:30Firstly, our previous studies have found that the weighted LDA based i-vector approach provides useful
0:01:37improvement over standard LDA based i-vector approach.
0:01:42However, there has been no detailed investigation on how weighted LDA dimension reduced with i-vector
0:01:48features with PLDA system, how it performs.
0:01:53In this paper, we hypothesized that weighted LDA and PLDA combined
0:01:59channel approach could do better job than existing approaches.
0:02:10In PLDA system we have been doing PLDA modeling and scoring on larger dimension and
0:02:13space, for example, five hundred.
0:02:16In dimension reduced PLDA system we have been doing the scoring and modeling on reduced
0:02:21base, hundred and hundred are the limited, so this technique considerably will reduce the computational
0:02:26complexity.
0:02:35Dimension reduced i-vector features based PLDA system.
0:02:45I-vector feature extractor already has been explained in previous presentation.
0:02:53The total variability space also.
0:02:56In tis section we have used pooled total variability approach for
0:03:03i-vector feature extraction.
0:03:09In this section I will talk about dimension reduced techniques
0:03:14the weighted LDA median fisher discriminant and weighted median fisher discriminant techniques.
0:03:23This is the version of approach which described how channel compensated i-vector features extractor.
0:03:32In the development phase channel compensated channel compensated i-vectors LDA, weighted LDA is median discriminant
0:03:36techniques
0:03:36are estimated in this following extractor
0:03:43After that, channel compensated i-vector features, w, have been
0:03:47estimated using channel processing.
0:04:02LDA followed by WCCN approach is commonly used in the various analyses of the i-vector
0:04:06system
0:04:10with PLDA system we've got PLDA composition.
0:04:14And now this is inaccurate. First stage, LDA it is based upon standard
0:04:20within class features that would
0:04:24p
0:04:25estimations
0:04:27and these are
0:04:32PLDA matrces are estimated using eigenvoices is b or sw.
0:04:42In the second stage, the WCCN is used to compensate
0:04:47everything WCCN is estimated based on estimating the matrix w
0:04:53and
0:04:57which represent
0:04:59finally, WCCN matrices are calculated using logs.
0:05:10Previously we have been standard LDA approach. Now we really opperate weighted LDA approach instead
0:05:17of standard LDA approach.
0:05:20In traditional LDA approach
0:05:23between class scatters don't take
0:05:26discriminative relationships between pairs of classes that are closer due to similarty. In this paper
0:05:30we have investigated weighted LDA. Weighting concepts are used in heavily weighted classes that are
0:05:37closer.
0:05:40The weighted between class scatter ... and these are already used in class-scatter relations.
0:05:56In this paper we investigated two different types of weighting functions. The first one is
0:06:00Euclidean distance weighting function.
0:06:06Second one is Mahalanobis
0:06:08distance weighting function.
0:06:10What that
0:06:14What are decreasing functions? And we're ginna analyze performances with different arbitrary values.
0:06:24All the weighted LDA techniques we calculated with weighted betweeen-class scatter, s b w.
0:06:34Weighted LDA matrix has similar
0:06:38standard LDA approach.
0:06:49Now we hear more on Median fisher discriminator. Previously, we discussed other LDA, weighted LDA,
0:06:55which is based on
0:06:57mean estimations.
0:06:59Median fisher discriminator between and within classs scatters can be estimated. The question arose why
0:07:08we have investigated median fisher discriminant analysis.
0:07:13In typical speaker verification system, we have only few recordings for each speaker. So averaging
0:07:18leads to loss of discriminant informatio
0:07:23Second one is
0:07:26median is used to estimate data with outliers.
0:07:37Median fisher discriminant algorithm
0:07:40Median based
0:07:41between and within class scatter estimations, using these approach, but here
0:07:53Average is calculated using
0:08:00Finally,
0:08:00median fisher discriminant matrix is calculated using eigenvector
0:08:15And PLDA approach and these were explained before two years.
0:08:19Presentation.
0:08:23But here, we have been doing PLDA modeling
0:08:37These were also explained before two years.
0:08:54Firstly,we have investigated LDA and weighted LDA approaches based on HTPLDA system.
0:09:00These were compared with standard HTPLDA system.
0:09:03can be also investigated Median fisher discriminator and weigh based HTPLDA system.
0:09:16Standard HTPLDA approach
0:09:24I-vector features think i
0:09:27UBM components and
0:09:30MFCC coeficients.
0:09:33The UBM was trained using these two thousand four telephone utterances.
0:09:38The total variability pooled weight, total variability approach PLDA,
0:09:43were trained using these two thousand four, two thousand six
0:09:47two thousand four two thousand five, six and Switchboard database.
0:09:53I-vectors were projected
0:09:57into LDA space using one hundred and fifty eigenvectors.
0:10:11Telephone and microphone pooled
0:10:12utterances form NIST two thousand four, two thousand five and six
0:10:15used for the score normalization.
0:10:28In the results and discussion section I will discuss
0:10:32between standard PLDA , the features as in HTPLDA system
0:10:43comparing the equal rate DCA performance within standard HTPLDA and LDA projected HTPLDA systems.
0:10:53Firstly, it can be clearly seen that LDA projected HTPLDA system
0:11:01perform better than standard HTPLDA system in microphone
0:11:10connected and weighted LDA, connected and weighted HTPLDA system
0:11:13projected HTPLDA system
0:11:16LDA projected HTPLDA system, all the conditions except
0:11:20telephone-telephone
0:11:32We have also investigated median fisher discriminator projected HTPLDA system
0:11:39and compared with standard HTPLDA system.
0:11:44For this case also
0:11:46we with HTPLDA system
0:11:54telephone condition.
0:11:57Median fisher discriminator
0:12:00improved equal rate performance in all the
0:12:05, across all the conditions.
0:12:19In pervious experiment we have found that LDA weighted-HTPLDA compared with weighted MFD with HTPLDA
0:12:22system show real improvement for my telephone conditions.
0:12:28The reason to keep that behaviour, telephone speakers i-vector discrimination is heavy-tailed.
0:12:36That's why we investigated median data discriminator, and weighted median fisher discriminator is good for
0:12:42data
0:12:47Compared all our system performance, standard HTPLDA and weighted
0:12:52LDA and HTPLDA system.
0:12:55Weighted median fisher discriminator with HTPLDA system.
0:13:09So, improvement on equal rate in telephone and telephone microphone speech.
0:13:16However, it doesn't
0:13:18improvement in DC of
0:13:32In this paper, we have investigated dimensionality techniques, such as LDA, weighted LDA,
0:13:37median fisher discriminator, weighted MDF with PLDA system.
0:13:44We have also found frome experiments that weighted LDA projected HTPLDA system has shown improvement
0:13:49in all conditions except telephone-telephone condi
0:13:53improvement in telephone conditions
0:14:00Weighted median fisher discriminator
0:14:03has shown as
0:14:05improvement at equal rate.
0:14:19Source normalized based LDA technique,
0:14:22normalized
0:14:23LDA technique,
0:14:27hasn't shown any major improvement
0:14:28over standard HTPLDA system.
0:14:34Source- normalized based and
0:14:36source-normalized weighted LDA techniques have shown major improvement on traditional i-vector based speaker verification sy
0:14:43Currently, these techniques
0:14:46are being investigated.
0:17:00Yeah, it was
0:17:02previously used
0:17:04Previously used by who?
0:17:09I've found some records. No.
0:17:35Well, my question is it obvious
0:17:37that using median base should perform better
0:17:48but I have studied the similar vectors
0:17:51and it so assimilates performance it doesn't make any improvement.
0:18:00We have tested median fisher discriminator technique and i-vector feature performances and the i-vector techniques.
0:18:06It doesn't show major improvement.
0:18:08that's only HTPLDA for dataset
0:20:12Yeah, that is to eliminate all of the directions taht are causing the problems in
0:20:17the microphone
0:21:23okay i think we could
0:21:27the speaker