0:00:15first i will give a quick overview of i-vectors
0:00:19after that i will
0:00:21only some of the methods for hand recounts and start the
0:00:26of the i-vector eyes them estimate scores by
0:00:30limited the
0:00:31duration of recordings
0:00:34then i will
0:00:36describe a simple preprocessing weighting scheme which uses duration information as a measure of
0:00:46i wrecked or a oral ability
0:00:49then i will describe some experiments and the results
0:00:54followed by concluding remarks
0:01:00in theory each decision should be made to dependent on the amount of data available
0:01:07and the same should hold also in the case of speaker recognition since
0:01:13we usually have recordings of different lengths
0:01:19in practice this is usually not the case mainly due to practical reasons since panic
0:01:27of uncertainty increases the article we agreed to make and computational complexity
0:01:34and also
0:01:36the gain in performance
0:01:38cohen
0:01:39can be not that could be not so significant especially if the recordings are sufficiently
0:01:46long
0:01:49in the case of
0:01:51i-vector challenge
0:01:53the
0:01:55i-vectors were extracted from recordings of different lengths
0:02:00and to the duration follows log normal distribution this suggests
0:02:06that
0:02:08we should see some improvement
0:02:11if the duration information is taken into account
0:02:18i-vector is defined as a map point estimate of keeping the variable of factor analysis
0:02:23model
0:02:24and it serves as a compact representation of speech utterance
0:02:31the posterior covariance encodes the answer t
0:02:36of the i-vector or
0:02:39estimate
0:02:42which is caused by a limited to duration of the recordings
0:02:46usually
0:02:47the i sort the
0:02:50is discarded to and comparing i-vectors for example in the
0:02:54be lda model
0:02:59nevertheless there have been proposed some solutions how to the
0:03:05take the uncertainty into account for example be a day with uncertainty propagation
0:03:11where and then we should note term is added to the model
0:03:16which models
0:03:17which explicitly models the
0:03:20duration variability
0:03:22another one
0:03:24is score calibration using different
0:03:26duration is a quality measure
0:03:28and yet another recycle i-vector scaling where the length normalisation is modified this to account
0:03:35for the
0:03:37uncertainty of i-vectors
0:03:43and those solutions are not directly applicable or at least not
0:03:48easily applicable in the context of i-vector challenge
0:03:53scenes
0:03:54the data for we can start
0:03:57reconstructing the posterior covariance is not available
0:04:01and also there is no development data that could be used for
0:04:08optimising the calibration parameters
0:04:12so is there another possibility how to use duration information
0:04:18to as a measure of i-vector a
0:04:22rely reliability
0:04:27prior to
0:04:32comparing the i-vectors are usually preprocessed
0:04:37among more common preprocessing methods are pca lda and do within class covariance normalization
0:04:46in which the basic step is to calculate mean and
0:04:52covariance matrix s
0:04:55we implicitly assume
0:04:56in those calculations that
0:04:59each the i-vector is equally all i-vectors are equally reliable
0:05:07some to account for the difference in a reliability of i-vectors
0:05:14re
0:05:14proposed a simple weighting scheme in of each other
0:05:21in which the to could contribution of each i-vector is multiplied by its corresponding duration
0:05:29so
0:05:30to verify the
0:05:34soundness of the proposed idea
0:05:36we implemented that the baseline system right in which we compare it
0:05:42the standard pca with
0:05:45the weighted version of the pca
0:05:49the results showed that weighted version of peace
0:05:52pca
0:05:56produce slightly better results than a standard one
0:06:01we also wanted to
0:06:04try within class covariance normalisation
0:06:07but
0:06:08in order to
0:06:10the apply within class covariance normalization
0:06:14we need to have labeled to date time which was not the case in the
0:06:19challenge
0:06:21so we needed to perform unsupervised the clustering we
0:06:28but
0:06:29experiment that with the different clustering algorithms but that the end to be selected k-means
0:06:35with cosine distance and four thousand clusters
0:06:42unfortunately the results are worse for within class covariance normalization then for a pca but
0:06:49at least the
0:06:51the weighted version was
0:06:54slightly a cat of the standard one
0:07:00we tried also several different classifiers and the best results were at used
0:07:07with a logistic regression but only after reading remove the
0:07:12length normalisation of from the processing pipeline
0:07:17in that case within class covariance normalisation
0:07:21gave better results then pca and all spend the can
0:07:27weighted towards and was
0:07:29score the better than standard one
0:07:35we try to further improve the results by additional fine-tuning
0:07:42so we put the duration as it and additional feature of i-vectors we excluded clusters
0:07:49with small official score
0:07:52we were is the roles of
0:07:56target and test i-vectors
0:07:59and do you can't do you want to the hyper parameters of logistic regression
0:08:04we did this fine tuning we were able to improve
0:08:09the previous result
0:08:11for a little bit more
0:08:13so this was also our
0:08:16best submitted result
0:08:20to conclude we
0:08:22present that
0:08:23a simple preprocessing weighting scheme which uses do duration information is a measure of i-vector
0:08:30a reliability
0:08:33we at you would quite reason the bus six sets
0:08:37with a clustering in the case of within class covariance normalization
0:08:42but okay but cat
0:08:45nearly no success with clustering in the case of the lda
0:08:50which suggests that we had a is more susceptible for labeling errors
0:08:56and the last remark we found out that length normalization does not help logistic regression
0:09:03thank you
0:09:21okay
0:09:31just empirical results but maybe somebody s can comment that i don't know
0:09:40nicole
0:09:46with on the side
0:09:47but we with at the same results as logistic regression icons otherwise
0:10:06did you generate what we did a clustering or you just one clustering stage we
0:10:10tried
0:10:11different things also two
0:10:14to iterate the clustering but didn't six it
0:10:26this was also experiments clear sets of four thousand because we didn't the get then
0:10:32you improvements by
0:10:35changing