Cochlear Mechanical Models Used in Automatic Speech Recognition Tasks

Abstract

In this paper we show that its possible unify two theories that we can find in the state of the art related with human hearing, one of them related with human perceptual phenomenon and the another one related with cochlear mechanic’s models linear. The first of them has been used since decade 1980’s into Automatic Speech Recognition Systems (ASRs) with satisfactory results. Whereas the second has been used since decade 1950’s but never used for ASRs. Since the second is the inner functionality with respect to the first, we propose that is very important to have a study about the behavior of the cochlea models into ASR tasks and compare the results that we can obtain. Then we present an auditory signal processing model that has been proposed as an alternative to the traditional filter banks and LPC models for speech spectral analysis. The argument for such a model is that, because it is based on known properties of the human auditory model (i.e. a model of the cochlea mechanics), it is inherently a better representation of the relevant spectral information that either a traditional bank-filter or an LPC model. In this work we use two different models of the cochlea that they are based in the classic mechanical to analyze their behavior when they are employed for ASR tasks with two variants and two more equations related with the place theory proposed by Von Bèkèsy. Also, we propose an alternative solution for another model based in the fluid mechanical. One time that we analyzed the response of the cochlea with different linear mechanical models we extracted features for ASR tasks that follow the cochlea behavior described by these models. The results obtained demonstrate that our propose represents a real alternative to be considered for this kind of computational applications. We obtained 2% of higher performance that when we used MFCC parameters in major cases.In this paper we show that its possible unify two theories that we can find in the state of the art related with human hearing, one of them related with human perceptual phenomenon and the another one related with cochlear mechanic’s models linear. The first of them has been used since decade 1980’s into Automatic Speech Recognition Systems (ASRs) with satisfactory results. Whereas the second has been used since decade 1950’s but never used for ASRs. Since the second is the inner functionality with respect to the first, we propose that is very important to have a study about the behavior of the cochlea models into ASR tasks and compare the results that we can obtain. Then we present an auditory signal processing model that has been proposed as an alternative to the traditional filter banks and LPC models for speech spectral analysis. The argument for such a model is that, because it is based on known properties of the human auditory model (i.e. a model of the cochlea mechanics), it is inherently a better representation of the relevant spectral information that either a traditional bank-filter or an LPC model. In this work we use two different models of the cochlea that they are based in the classic mechanical to analyze their behavior when they are employed for ASR tasks with two variants and two more equations related with the place theory proposed by Von Bèkèsy. Also, we propose an alternative solution for another model based in the fluid mechanical. One time that we analyzed the response of the cochlea with different linear mechanical models we extracted features for ASR tasks that follow the cochlea behavior described by these models. The results obtained demonstrate that our propose represents a real alternative to be considered for this kind of computational applications. We obtained 2% of higher performance that when we used MFCC parameters in major cases.