| 0:00:16 | so |
|---|
| 0:00:18 | i'm going to have proposed here with my only collectively from if you k |
|---|
| 0:00:23 | it's rental |
|---|
| 0:00:24 | and basically score multichannel i-vector combination for robust speaker verification from the most environments |
|---|
| 0:00:32 | basically we were well a couple of years ago in a in a project about |
|---|
| 0:00:35 | a automation application basically was glued inner product and the main characteristics of the project |
|---|
| 0:00:44 | was uttered below a kind of or |
|---|
| 0:00:47 | of |
|---|
| 0:00:49 | speech interface in order to talk to activate the windows the doors and so on |
|---|
| 0:00:54 | for people with physical impairments |
|---|
| 0:00:56 | i either they're |
|---|
| 0:00:58 | them as you mentioned above the main that this is was that the use or |
|---|
| 0:01:01 | try to users like a multiple microphones are not was from the big microphones with |
|---|
| 0:01:06 | of the system always listening and of course the use was supposed to be able |
|---|
| 0:01:09 | to say comments from anywhere in any position in the room |
|---|
| 0:01:14 | so it we of course that some speaker tasks and some speaker services we were |
|---|
| 0:01:21 | there's the in and basically the conference we were expecting in this kind of applications |
|---|
| 0:01:29 | not it was |
|---|
| 0:01:30 | of course of the speaker can be anyone so we expect to have a huge |
|---|
| 0:01:34 | mismatch of in between enrollment the spaces |
|---|
| 0:01:37 | even if you have fixed the problem and or even clean enrolment with a and |
|---|
| 0:01:42 | i from application about as more from location |
|---|
| 0:01:44 | and of course we need to the models that were able to cope with this |
|---|
| 0:01:47 | problem a typically wouldn't five number of microphones the solution that |
|---|
| 0:01:52 | you can find |
|---|
| 0:01:54 | in the literature are trying to play something of the speech enhancement even microphone array |
|---|
| 0:02:00 | beamforming and someone usually you need but it will clearly that devices for that and |
|---|
| 0:02:05 | sensors you can do some common combination the post processed with the to describe it |
|---|
| 0:02:10 | some |
|---|
| 0:02:12 | and to play every with i-vectors or to combine different channels in the sense different |
|---|
| 0:02:17 | channels that are recorded in a at the same utterance so we have |
|---|
| 0:02:22 | the same |
|---|
| 0:02:24 | different samples of the same utterance i for that we would basically use this apartment |
|---|
| 0:02:29 | with we may sure many impulse responses with a from many locations of positions and |
|---|
| 0:02:35 | which relate the database |
|---|
| 0:02:37 | i'm basically we will that the scenario we will focus on some have been a |
|---|
| 0:02:41 | fixed position for enrollment and you can be anywhere for speaker verification we are explained |
|---|
| 0:02:48 | every for the combined effect but we can try and we can talk more an |
|---|
| 0:02:52 | impostor |
|---|