| Abstract: |
A new approach to speaker verification (SV) is developed, which decides for two given speech signals, whether they have been spoken by the same person or not. The idea is to apply pattern
matching for two speech signals, which are not exactly equally worded, but contain equally worded segments. This is done by first searching equally worded segments which are subsequently used for
pattern matching and finally for taking the decision. For the DTW-based pattern matching, instead of an Euclidean distance measure the probability that the corresponding frames are from the same
speaker is used. This probability is determined by an appropriately trained neural network. This approach to SV is completely language-independent, to some degree text-independent and it needs no
speaker enrollment. |
