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