| 0:00:15 | i don't |
|---|
| 0:00:18 | i'm going to percent for about the speaker recognition in formant frequencies in linguistic units |
|---|
| 0:00:24 | and the motivation is the wavelet use of the formant frequencies applied in linguistic constraints |
|---|
| 0:00:30 | in four and six but there's a need to validate these formant based speaker discrimination |
|---|
| 0:00:37 | of formant frequencies from the standard benchmark by nist sre |
|---|
| 0:00:44 | so you know previous work we present an approach to in which we extract i-vectors |
|---|
| 0:00:51 | from the segments belonging to a specific and linguistic constraints |
|---|
| 0:00:57 | okay |
|---|
| 0:00:59 | based on formant frequencies and we use the score a cosine scoring and score normalization |
|---|
| 0:01:05 | and not clear is understanding of the to obtain a wellcalibrated slant ranges better linguistic |
|---|
| 0:01:13 | study |
|---|
| 0:01:14 | i in this work for a sequence as existing we replace a cosine scoring |
|---|
| 0:01:22 | and score normalization and calibration steps |
|---|
| 0:01:26 | with a covariance model |
|---|
| 0:01:28 | and based on the same linguistically constraining formation i-vectors |
|---|
| 0:01:34 | and be used in a improve discrimination which is not surprising |
|---|
| 0:01:39 | but the thing is that we obtain a scores with low below countries you're a |
|---|
| 0:01:44 | loss |
|---|
| 0:01:45 | the constraint so we can be used directly as like radios |
|---|
| 0:01:50 | a million the an additional information from score school |
|---|
| 0:01:55 | like right yes |
|---|
| 0:01:57 | bizarre the results when combining several of these linguistic constraints |
|---|
| 0:02:03 | on nist sre two thousand and six |
|---|
| 0:02:08 | decide there is also just remind that |
|---|
| 0:02:11 | it and the results are using a and formant frequencies |
|---|
| 0:02:17 | that's this summer i u one two normal at a site would be of course |
|---|
| 0:02:20 | that's ser mean of stardom or think |
|---|