Requency-distance Cochlea Equations Applied in Automatic Recognition Audio Signals
Abstract
This paper shows the study of two methodologies based on physiological models of the cochlea for the analysis of speech, music and other acoustic signals. The above with the objective to find equations that describe the distance where the cochlea is excited when a determined frequency is produced and perceived by the human ear according to the physiological models used. It’s very important to mention that cochlea behavior must to be studied to analyze how to construct new parameters to be used in Automatic Recognition in Speech Recognition (ASR) and Music Transcription (MT). In this paper we use two cochlear models to probe how it’s possible to find a set of parameters to be employed for the tasks to analyze and recognize of the audio signals. To obtain these set of parameters different computational algorithms were used, to mention some of them: electrical network solutions by partial fractions or successive approximations, resonance analysis, non-linear least squares; among others. The objective, independently of the cochlear model employed, it was to find an equation that related frequency values vs distance that describe the behavior of the cochlea. In consequence, different parameters that describe the behavior of the cochlea were used for Music Transcription (MT) and Automatic Speech Recognition (ASR) tasks. After that, our propose was compared with another equation founded in state of the art developed by Greenwood and we analyzed the results and difference between them. We obtained better results with our purpose in comparison with another used.
Keywords
Cochlear mechanics model, non-linear regression, frequency-position function, music transcription and automatic speech recognition