Noise removal feature enhancement and speech recognition techniques for artificial larynx transducer speech

Abstract

Speech impediments are the state of difficulty for a person to speak comfortably. These impediments make the spoken speech distorted and they are generally categorized as disordered speech. The quality of disordered speech is poor as clarity, intelligibility and naturalness is missing. In most type of disordered speech the voice is natural and produced by the vocal system of the human being. The vocal system includes the organ called as Larynx placed in the upper part of the neck. This organ has the vocal folds that contribute for pitch variation and volume of the speech. This organ will be malfunctioning some time or will be removed because of cancer. In both the case in order to restore speech, an external device called Artificial Larynx Transducer (ALT) is used to produce the sound. It is a small handheld battery operated device and is used for decades to obtain the audible speech for people who lost their speech because of removal of larynx. The quality of speech and its intelligibility of AL speakers have not improved for decades. The reason for poor quality is constant vibration of ALT, direct sound from ALT and pressure offered to produce the vibration. newlineSo in this research the nature of the speech produced from ALT is analyzed, a possible enhancement of the parameter is done and a recognition technique of the spoken word with the help of trained data is done. Here the approach followed to tackle the problem of poor quality in AL speech involves both speech enhancement and recognizer technique development. When it is looked as enhancement problem noise region localization, noise estimation and noise suppression methods were adopted. In the process of parameter enhancement, pitch frequency estimation and improvement is implemented. When it is looked as recognition problem the parameters pitch frequency, formant frequency, glottal excitation, spectral tilt, coefficients are extracted. As formant frequency is a sensitive parameter, its estimation was done using Recurrent Neural network.