Formant Frequency Estimation of Artificial Larynx Transducer Speech Using Recurrent Neural Network

K. Inbanila and E. Krishna Kumar

Human Beings communicate with each other by speaking. The speech as a signal has 2 components voiced and unvoiced speech. Voiced speech is produced by the excitation produced at glottis and unvoiced is pro-duced by noise created at the mouth. The voiced components that is produced at glottis passes through the vocal tract and then reach the mouth. The nature of the speech is determined mostly at the vocal tract. But for some reason for some people the speech produced is not proper because of the organ problems or motor disorder issue. In these cases, the speech produced is called disordered speech and termed with the names like stammering, apraxia, dysar-theria and so on. In some case, the larynx is removed from human body because of cancer or other issues. For them, Artificial Larynx Transducer (ALT) is given to produce substitution voice. This paper aims at formant frequency estimation of the speech produced with the help of ALT using Recurrent Neural Network (RNN) method. The speech produced with the help of ALT will lack in naturalness and intelligibility. The direct noise coming from ALT device is called DREL noise. This also creates irritability to the listener. So in this paper, a method is proposed for the DREL noise removal and formant frequency estimation of the ALT speech

Volume 11 | 01-Special Issue

Pages: 72-81