Reconstruction of Dysphonic Speech by MELP

Abstract

The chronical dysphony is the result of neural, structural or phathological effects on the vocal cords or larynx and it causes undesirable changes in the quality of speech. This paper presents a Mixed Excitation Linear Prediction (MELP) based system that reconstructs normally phonated speech from dysphonic speech, while preserving the individuality of the patient. The proposed system can be used as speech prosthesis for the patients who have lost the ability to produce voice. To reconstruct normally phonated speech from dysphonic speech, pitch generation using the perceived pitch relationship with formant frequencies, formant and voicing modification steps were performed for phonemes. The principle novelty of this study is to modify voiced phonemes' acoustic features while preserving unvoiced ones. Therefore voiced-unvoiced detection is performed for each phoneme. The proposed system is composed of three main parts. In the analysis phase the acoustic differences observed between normal and dysphonic speech are determined. Acoustic parameters of the dysphonic speech's voiced phonemes are modified in order to obtain a synthetic speech that is closer to normal speech. Finally, enhanced speech is synthesized by MELP.