| 0:00:15 | an initial draft and that would like to well |
|---|
| 0:00:18 | a given introduction to the to the work done by olivier models on all of |
|---|
| 0:00:22 | the s two cannot be here |
|---|
| 0:00:25 | and try to be quick |
|---|
| 0:00:28 | so of the work is about some analysis homework a sorry forgot the title skull |
|---|
| 0:00:33 | analysis an optimization of bottleneck features for speaker recognition |
|---|
| 0:00:37 | and it's basically are about features that are used as an input to the to |
|---|
| 0:00:42 | the widely used bottleneck features |
|---|
| 0:00:47 | so first we started by using the of the sr features and try to optimize |
|---|
| 0:00:52 | the network to give the best phoneme recognition for phoneme error rate basically |
|---|
| 0:00:58 | and what we see that it is that she try to put in the features |
|---|
| 0:01:02 | that we know there are best for these speaker recognition which is the mfccs plus |
|---|
| 0:01:07 | some extra normalisation plus some extra constraint on the on the neural network training et |
|---|
| 0:01:14 | cetera so the conclusion is that the asr features |
|---|
| 0:01:18 | are not the best for of speaker recognition using the bottleneck ubm i-vector system |
|---|
| 0:01:24 | and better phone accuracy does not necessarily mean a better speaker error rates so thank |
|---|
| 0:01:31 | you once the poster |
|---|