It has been shown in several recent publications that application of vocal tract normalization (VTN) is a successful method for improving the accuracy of speaker independent recognisers. We argue that VTN can be implemented in the filterbank domain and propose a model to achieve this. We show how the model can be implemented directly in the MFCC domain, where it may be viewed as a constrained version of maximum likelihood linear regression (MLLR). The parameter estimates produced by the model are in accord with our ideas about how it should operate to perform VTN. Recognition results on a phoneme recognition task are presented which show a small improvement in accuracy.
|Number of pages||4|
|Publication status||Published - Oct 2000|
|Event||Sixth International Conference on Spoken Language Processing (ICSLP 2000) - Beijing, China|
Duration: 16 Oct 2000 → 20 Oct 2000
|Conference||Sixth International Conference on Spoken Language Processing (ICSLP 2000)|
|Period||16/10/00 → 20/10/00|