An Adaptive Q Cochlear Filter in Phoneme Recognition

Abstract

Introduction It has been expected that speech recognition performance can be improved by replacing a traditional front-end with a model of the auditory periphery. The underlying assumption is that if a model could be designed properly, it should generate a more efficient representation compared to traditional physical spectrum representations. From this viewpoint, several works have been reported [11-[41. However, these studies do not always show an auditory front-end to be superior to a traditional front-end. Some auditory front-ends are superior only for noisy speech, but many show little, if any, superiority in processing clean speech. We also have been developing an auditory model characterized by an adaptive Q cochlear filter not only for the front-end of a speech recognition system but also for a general purpose spectral1 analyzer in speech research [SI. In this paper, several auditory front-ends based on the adaptive Q cochlear filter and its relatives are tested in speaker dependent phoneme recognition using different stochastic pattern classifier!;, a shift invariant multi template matching system using LVQ2-trained codebook, and a VQ-HMM system. Further, we will discuss problems of using an auditory model as a frontend of an automatic speech recognition system. 2. Adaptive Q Cochlear Filter An adaptive Q cochlear filter (AQF) is a computational filter that functionally simulates the nonlinear filtering characteristics of the basilar membrane vibrating system. The AQF consists of three parts: (1) cascaded second-order notch filters (NOTCH), (2) second-order band pass filters (BPF) connected to each NOTCH output, and (3) adaptive Q circuits connected to each BPF output. The adaptive Q circuit consists of a second-order low-pass filter (LPF) in