0:00:19 | PLDA based speaker verification with weighted LDA do techniques. |
---|---|

0:00:28 | This is the outline of my presentation. First part is motivation, where I will discuss |

0:00:30 | why we have investigated different techniques such as LDA weighted with PLDA system. |

0:00:43 | Using dimensionality reduction on i-vector features. |

0:00:48 | I will then discuss the experiments on telephone and microphone speech with PLDA system |

0:00:54 | which is based on LDA and weighted LDA dimension reduction techniques. |

0:01:03 | Our main motivation in this paper is to identify the best channel compensation approach for |

0:01:06 | telephone and microphone based speaker verification system. |

0:01:12 | Dehak has investigated dimensionality reduction techniques for channel compensation is the i-vector system. |

0:01:19 | And he has investigated PLDA modeling with i-vectors to compensate channel variability. |

0:01:30 | Firstly, our previous studies have found that the weighted LDA based i-vector approach provides useful |

0:01:37 | improvement over standard LDA based i-vector approach. |

0:01:42 | However, there has been no detailed investigation on how weighted LDA dimension reduced with i-vector |

0:01:48 | features with PLDA system, how it performs. |

0:01:53 | In this paper, we hypothesized that weighted LDA and PLDA combined |

0:01:59 | channel approach could do better job than existing approaches. |

0:02:10 | In PLDA system we have been doing PLDA modeling and scoring on larger dimension and |

0:02:13 | space, for example, five hundred. |

0:02:16 | In dimension reduced PLDA system we have been doing the scoring and modeling on reduced |

0:02:21 | base, hundred and hundred are the limited, so this technique considerably will reduce the computational |

0:02:26 | complexity. |

0:02:35 | Dimension reduced i-vector features based PLDA system. |

0:02:45 | I-vector feature extractor already has been explained in previous presentation. |

0:02:53 | The total variability space also. |

0:02:56 | In tis section we have used pooled total variability approach for |

0:03:03 | i-vector feature extraction. |

0:03:09 | In this section I will talk about dimension reduced techniques |

0:03:14 | the weighted LDA median fisher discriminant and weighted median fisher discriminant techniques. |

0:03:23 | This is the version of approach which described how channel compensated i-vector features extractor. |

0:03:32 | In the development phase channel compensated channel compensated i-vectors LDA, weighted LDA is median discriminant |

0:03:36 | techniques |

0:03:36 | are estimated in this following extractor |

0:03:43 | After that, channel compensated i-vector features, w, have been |

0:03:47 | estimated using channel processing. |

0:04:02 | LDA followed by WCCN approach is commonly used in the various analyses of the i-vector |

0:04:06 | system |

0:04:10 | with PLDA system we've got PLDA composition. |

0:04:14 | And now this is inaccurate. First stage, LDA it is based upon standard |

0:04:20 | within class features that would |

0:04:24 | p |

0:04:25 | estimations |

0:04:27 | and these are |

0:04:32 | PLDA matrces are estimated using eigenvoices is b or sw. |

0:04:42 | In the second stage, the WCCN is used to compensate |

0:04:47 | everything WCCN is estimated based on estimating the matrix w |

0:04:53 | and |

0:04:57 | which represent |

0:04:59 | finally, WCCN matrices are calculated using logs. |

0:05:10 | Previously we have been standard LDA approach. Now we really opperate weighted LDA approach instead |

0:05:17 | of standard LDA approach. |

0:05:20 | In traditional LDA approach |

0:05:23 | between class scatters don't take |

0:05:26 | discriminative relationships between pairs of classes that are closer due to similarty. In this paper |

0:05:30 | we have investigated weighted LDA. Weighting concepts are used in heavily weighted classes that are |

0:05:37 | closer. |

0:05:40 | The weighted between class scatter ... and these are already used in class-scatter relations. |

0:05:56 | In this paper we investigated two different types of weighting functions. The first one is |

0:06:00 | Euclidean distance weighting function. |

0:06:06 | Second one is Mahalanobis |

0:06:08 | distance weighting function. |

0:06:10 | What that |

0:06:14 | What are decreasing functions? And we're ginna analyze performances with different arbitrary values. |

0:06:24 | All the weighted LDA techniques we calculated with weighted betweeen-class scatter, s b w. |

0:06:34 | Weighted LDA matrix has similar |

0:06:38 | standard LDA approach. |

0:06:49 | Now we hear more on Median fisher discriminator. Previously, we discussed other LDA, weighted LDA, |

0:06:55 | which is based on |

0:06:57 | mean estimations. |

0:06:59 | Median fisher discriminator between and within classs scatters can be estimated. The question arose why |

0:07:08 | we have investigated median fisher discriminant analysis. |

0:07:13 | In typical speaker verification system, we have only few recordings for each speaker. So averaging |

0:07:18 | leads to loss of discriminant informatio |

0:07:23 | Second one is |

0:07:26 | median is used to estimate data with outliers. |

0:07:37 | Median fisher discriminant algorithm |

0:07:40 | Median based |

0:07:41 | between and within class scatter estimations, using these approach, but here |

0:07:53 | Average is calculated using |

0:08:00 | Finally, |

0:08:00 | median fisher discriminant matrix is calculated using eigenvector |

0:08:15 | And PLDA approach and these were explained before two years. |

0:08:19 | Presentation. |

0:08:23 | But here, we have been doing PLDA modeling |

0:08:37 | These were also explained before two years. |

0:08:54 | Firstly,we have investigated LDA and weighted LDA approaches based on HTPLDA system. |

0:09:00 | These were compared with standard HTPLDA system. |

0:09:03 | can be also investigated Median fisher discriminator and weigh based HTPLDA system. |

0:09:16 | Standard HTPLDA approach |

0:09:24 | I-vector features think i |

0:09:27 | UBM components and |

0:09:30 | MFCC coeficients. |

0:09:33 | The UBM was trained using these two thousand four telephone utterances. |

0:09:38 | The total variability pooled weight, total variability approach PLDA, |

0:09:43 | were trained using these two thousand four, two thousand six |

0:09:47 | two thousand four two thousand five, six and Switchboard database. |

0:09:53 | I-vectors were projected |

0:09:57 | into LDA space using one hundred and fifty eigenvectors. |

0:10:11 | Telephone and microphone pooled |

0:10:12 | utterances form NIST two thousand four, two thousand five and six |

0:10:15 | used for the score normalization. |

0:10:28 | In the results and discussion section I will discuss |

0:10:32 | between standard PLDA , the features as in HTPLDA system |

0:10:43 | comparing the equal rate DCA performance within standard HTPLDA and LDA projected HTPLDA systems. |

0:10:53 | Firstly, it can be clearly seen that LDA projected HTPLDA system |

0:11:01 | perform better than standard HTPLDA system in microphone |

0:11:10 | connected and weighted LDA, connected and weighted HTPLDA system |

0:11:13 | projected HTPLDA system |

0:11:16 | LDA projected HTPLDA system, all the conditions except |

0:11:20 | telephone-telephone |

0:11:32 | We have also investigated median fisher discriminator projected HTPLDA system |

0:11:39 | and compared with standard HTPLDA system. |

0:11:44 | For this case also |

0:11:46 | we with HTPLDA system |

0:11:54 | telephone condition. |

0:11:57 | Median fisher discriminator |

0:12:00 | improved equal rate performance in all the |

0:12:05 | , across all the conditions. |

0:12:19 | In pervious experiment we have found that LDA weighted-HTPLDA compared with weighted MFD with HTPLDA |

0:12:22 | system show real improvement for my telephone conditions. |

0:12:28 | The reason to keep that behaviour, telephone speakers i-vector discrimination is heavy-tailed. |

0:12:36 | That's why we investigated median data discriminator, and weighted median fisher discriminator is good for |

0:12:42 | data |

0:12:47 | Compared all our system performance, standard HTPLDA and weighted |

0:12:52 | LDA and HTPLDA system. |

0:12:55 | Weighted median fisher discriminator with HTPLDA system. |

0:13:09 | So, improvement on equal rate in telephone and telephone microphone speech. |

0:13:16 | However, it doesn't |

0:13:18 | improvement in DC of |

0:13:32 | In this paper, we have investigated dimensionality techniques, such as LDA, weighted LDA, |

0:13:37 | median fisher discriminator, weighted MDF with PLDA system. |

0:13:44 | We have also found frome experiments that weighted LDA projected HTPLDA system has shown improvement |

0:13:49 | in all conditions except telephone-telephone condi |

0:13:53 | improvement in telephone conditions |

0:14:00 | Weighted median fisher discriminator |

0:14:03 | has shown as |

0:14:05 | improvement at equal rate. |

0:14:19 | Source normalized based LDA technique, |

0:14:22 | normalized |

0:14:23 | LDA technique, |

0:14:27 | hasn't shown any major improvement |

0:14:28 | over standard HTPLDA system. |

0:14:34 | Source- normalized based and |

0:14:36 | source-normalized weighted LDA techniques have shown major improvement on traditional i-vector based speaker verification sy |

0:14:43 | Currently, these techniques |

0:14:46 | are being investigated. |

0:17:00 | Yeah, it was |

0:17:02 | previously used |

0:17:04 | Previously used by who? |

0:17:09 | I've found some records. No. |

0:17:35 | Well, my question is it obvious |

0:17:37 | that using median base should perform better |

0:17:48 | but I have studied the similar vectors |

0:17:51 | and it so assimilates performance it doesn't make any improvement. |

0:18:00 | We have tested median fisher discriminator technique and i-vector feature performances and the i-vector techniques. |

0:18:06 | It doesn't show major improvement. |

0:18:08 | that's only HTPLDA for dataset |

0:20:12 | Yeah, that is to eliminate all of the directions taht are causing the problems in |

0:20:17 | the microphone |

0:21:23 | okay i think we could |

0:21:27 | the speaker |