A neural network model of hidden markov model applied to the auditory periphery for speech processing and recognition

This paper presents a neural network model called Hidden Markov Model (HMM). It is applied to analysis and recognize the response of the auditory periphery to speech stimulation. Based on the reports from some scientists on the world, the electric discharge rate of auditory nerve and the otoacoustic emissions of auditory periphery underlying speech stimulation are related. This relation supports the idea that the response of auditory periphery to speech is a random procedure of double layers. For the response of a short-time, the procedure is stationary time-invariant. For a long sequence response, the response procedure consists of many short-time states and the transition from one state to next state is governed by a set of transition probabilities. The procedure of recognition is to know which observation vector is more matched with the codebook. Codebook is governed by a set of output probabilities. Hidden Markov Model just is an appropriate algorithm to represent the random procedure of double layers. Moreover, in order to reduce the number of input parameters of HMM neural network, the vector quantization (VQ) is applied to converse the characteristic vectors to the observation values. For our experiment, a programmable device used to measure and process the response of auditory periphery is developed in our laboratory. The device not only can detect transient otoacoustic emissions(TEOAE) and the distortion product otoacoustic (DPOAE), but also can synthesize any stimulation, such as speech, and then receive generated response in auditory periphery, finally automated recognize the response. The prime results presented in the paper show that HMM has some of potential possibilities in the applications of speech processing and recognition of auditory periphery. Therefore, the further research will benefit to design hearing aids and a front-end for speech recognition, etc.