Reconstruction of Dysphonic Speech for Synthesizing Normally Phonated Speech

Abstract

In this study, a novel system, delivering synthetic speech with the quality near to natural, is designed and implemented by reconstructing dysphonic speech of patients that have lost their voice totally due to apoplectic chordae vocalis, organic lesions of vocal cords or partial laryngectomy. In the proposed system, A MELP (Mixed Excitation Linear Prediction) is used for synthesizing the normal speech. The unvoiced phonemes are determined in dysphonic speech and synthetic pitch is generated by using pitch-formant frequency relation and then formant distortion modification is applied and voicing is added for the phonemes other than unvoiced phonemes. MELP will be used for synthesizing enhanced speech by using modified acoustic features. Spectral distance measurement is made and subjective listening tests are applied for assessing the produced synthetic speech by our proposed system. Our tests show that, synthetic speech produced this way is preferable when compared to dysphonic speech especially in terms of timbre, recognition and naturality.