| 0:00:15 | although an afternoon |
|---|
| 0:00:18 | it's a problem i mean every to mister technology later to present my paper this |
|---|
| 0:00:23 | conference |
|---|
| 0:00:26 | so what the problem is a sampling to address the problem of variability in the |
|---|
| 0:00:31 | i-vector space you to the acoustic content of the |
|---|
| 0:00:36 | speech |
|---|
| 0:00:37 | and the language is the main one of the main now |
|---|
| 0:00:40 | source of this variability |
|---|
| 0:00:43 | the probabilistic linear discriminant analysis while going to model the kind of source variability but |
|---|
| 0:00:51 | it cannot some model this variability using multilingual without a multilingual apartments that's for each |
|---|
| 0:01:00 | speaker |
|---|
| 0:01:01 | so |
|---|
| 0:01:03 | there is a one a method called language normalized w c n |
|---|
| 0:01:08 | which is designed to model this variability by extending the source normalized mfcc |
|---|
| 0:01:16 | no |
|---|
| 0:01:17 | this is done before a prior to the p lda training camp |
|---|
| 0:01:21 | so what i am going to post is to |
|---|
| 0:01:26 | propose a purely training algorithm |
|---|
| 0:01:29 | we would be built to reduce this language in fact |
|---|
| 0:01:33 | so by estimating the speaker and channel subspace stuff from multilingual utterances your we can |
|---|
| 0:01:40 | be appealed it can be able to work independent |
|---|
| 0:01:47 | so when evaluated on the nist sre two thousand eight core condition |
|---|
| 0:01:53 | we were able to you know |
|---|
| 0:01:57 | reduce the fact |
|---|
| 0:01:59 | we use the russian spanish arabic and mandarin in addition to english |
|---|
| 0:02:04 | utterances |
|---|
| 0:02:07 | so in comparison with the baseline system we use double system was we were able |
|---|
| 0:02:13 | to |
|---|
| 0:02:14 | we choose the language effects by ten percent the equal error rates |
|---|
| 0:02:22 | so that's it |
|---|
| 0:02:24 | okay |
|---|