Abstract
A new dynamic weighting method for robust text-dependent speaker verification is proposed and investigated. The main attraction of the proposed approach is that it is equally effective under both uniform and non-uniform mismatched conditions. It involves estimating the mismatch associated with each feature vector in the test token and using this to weight the respective vector distortion appropriately, prior to the computation of the overall degree of dissimilarity. The experiments were conducted using a subset of the Brent speech database consisting of repetitions of isolated digit utterances zero to nine spoken by native English speakers. Based on the experimental results it is shown that the use of the proposed approach leads to a considerably higher accuracy in speaker verification than that obtainable with conventional score normalisation methods.
| Original language | English |
|---|---|
| Title of host publication | 6th European Conference on Speech Communication and Technology (Eurospeech '99) |
| Pages | 967-970 |
| Number of pages | 4 |
| Publication status | Published - 1999 |