Improved data modeling for text-dependent speaker recognition using sub-band processing

A growing body of recent work documents the potential benefits of sub-band processing over wideband processing in automatic speech recognition and, less usually, speaker recognition. It is often found that the sub-band approach delivers performance improvements (especially in the presence of noise), but not always so. This raises the question of precisely when and how sub-band processing might be advantageous, which is difficult to answer because there is as yet only a rudimentary theoretical framework guiding this work. We describe a simple sub-band speaker recognition system designed to facilitate experimentation aimed at increasing understanding of the approach. This splits the time-domain speech signal into 16 sub-bands using a bank of second-order filters spaced on the psychophysical mel scale. Each sub-band has its own separate cepstral-based recognition system, the outputs of which are combined using the sum rule to produce a final decision. We find that sub-band processing leads to worthwhile reductions in both the verification and identification error rates relative to the wideband system. decreasing the identification error rate from 3.33% to 0.56% and equal error rate for verification by approximately 50% for clean speech. The hypothesis is advanced that, unlike the wideband system, sub-band processing effectively constrains the free parameters of the speaker models to be more uniformly deployed across frequency: as such, it offers a practical solution to the bias/variance dilemma of data modeling. Much remains to be done to explore fully the new paradigm of sub-band processing. Accordingly, several avenues for future are identified. In particular, we aim to explore the hypothesis of a practical solution to the bias/variance dilemma in more depth.