| 0:00:15 | thank you a this work was a menu that done by yourself this of the |
|---|
| 0:00:18 | rich from these ridiculous but you couldn't travels or would be presenting |
|---|
| 0:00:23 | the goal here is basically to deal with a the problem of calibration or score |
|---|
| 0:00:28 | normalization and the process of a noisy speech |
|---|
| 0:00:31 | and |
|---|
| 0:00:32 | the framework the speaker pollution framework is i-vector a with cosine distance |
|---|
| 0:00:38 | so what we proposed to do is to estimate a noise i-vector using nonspeech portions |
|---|
| 0:00:44 | of the signal and use this to predict |
|---|
| 0:00:47 | the noise impacts on the i-vector space and on the score |
|---|
| 0:00:55 | so basically if we define x as the i-vector for clean speech with which is |
|---|
| 0:00:59 | not served and used i-vector from nonspeech we can estimate form |
|---|
| 0:01:03 | noisy portions and iced i-vector for noisy or observed speech |
|---|
| 0:01:08 | i'll five is directly off a standard deviation of the most signal in noisy speech |
|---|
| 0:01:13 | signal so it's related to snr so basically if for clean speech also is zero |
|---|
| 0:01:18 | and da actually the |
|---|
| 0:01:21 | the observed a an i-vector is actually equal to what we want to actually have |
|---|
| 0:01:26 | dark clean i-vector |
|---|
| 0:01:27 | and for extremely noisy is a speech we only observe that the noise what we |
|---|
| 0:01:32 | can try to due to the some a linear approximation |
|---|
| 0:01:36 | and use this to estimate it down the bias |
|---|
| 0:01:40 | a bias terminus scaling the term for a scoring function and use this vocal which |
|---|
| 0:01:45 | thank you |
|---|