5th International Conference on Spoken Language Processing
The concept of splitting the entire frequency domain into sub-bands and processing the spectra in these bands independently in between every consecutive recombination stage to generate a final score has recently been proposed for speech recognition. Some of the aspects of this technique have also been studied for the task of speaker recognition. A remaining critical problem in this approach, however, is the determination of appropriate recombination weights. This paper presents a new method for generating these weights for sub-band based speaker verification. The approach is based on the use of background speaker models and aims to reduce the effect of any mismatch between the band-limited segments of the test utterance and the corresponding sections in the target speaker model. The paper also discusses a problem generally associated with the sub-band cepstral features and outlines a possible solution.
Bibliographic reference. Sivakumaran, Perasiriyan / Ariyaeeinia, Aladdin M. / Hewitt, Jill A. (1998): "Sub-band based speaker verification using dynamic recombination weights", In ICSLP-1998, paper 1055.