Multiple sub-band systems for speaker verification

It is well known that the spectral information contained in the full-band cepstral parameters within a certain range is highly useful for speaker discrimination. However, this information cannot be fully represented by using the cepstral parameters generated in a single sub-band-system. This paper focuses on methods to tackle this deficiency of the sub-band cepstum in the context of text-dependent speaker verification. In particular, it focuses on a technique based on the collective use of the cepstral vectors generated from a set of different sub-band systems. The paper also addresses the procedure for an effective incorporation of this technique into the hidden Markov model (HMM) framework. Furthermore, the implementation issues are discussed and details of the experimental evaluation are presented.