Speaker identification using subband HMMS

This paper is concerned with optimum band splitting and optimum recombination weights in subband HMM-based speaker identication. In the rst experiment , the full frequency band (8kHz) was split into two subbands, and speaker identication rate was measured for various splitting frequencies and recom-bination weights. It was found that subbands 0-2kHz and 2-8kHz with equal recombination weights gave the best identication rate, outperforming a base-line method without band-splitting. In the second experiment, the full-band was split into three sub-bands with various splitting frequencies. Splitting into 0-2kHz, 2-6kHz, and 6-8kHz gave the best result, slightly outperforming the two-subband case. Finally, four-subband experiment was conducted, the result of which suggests that the speaker information and the phonemic information are complementary to a considerable degree in the spectral domain.