Reconstruction Of Normal Speech From Chinese Whispers Based On MELP

As an unusual modality of speech communication, whispered speech has received growing attention from professionals in the fields of phonetics and iatrology in recent years. Reconstruction of normal speech from whispered speech can be applied in several fields, such as the private speech communication using cell-phone in public, the assistant to natural communication for the individuals with disease of larynx, the special need for the forensic work, etc. The research of the reconstruction has significant academic and practical value.The realization of reconstructing from Chinese whispers is mainly discussed in this thesis. A reconstruction system based on the mixed excitation linear prediction model (MELP) is proposed. According to the principal of MELP coder, the necessary parameters of normal speech are extracted and then used to synthesize speech. A MELP synthesis system is primarily founded. Then, according to the MELP synthesis system and the differences between Chinese whispers and normal speech, the parameters of whispers are analyzed, compared and then modified to the parameters which are needed in reconstructing from whispers. In this thesis, a method of endpoint detection of Chinese whispers based on order statistic filter (OSF) is proposed, a method of differentiating speakers is introduced and a regression model of transforming the value of gain of whispers to that of normal speech based on statistic is indicated. The problems of initial/final detection, the identification of tones and the modification of formants in Chinese whispers are briefly introduced. A method of estimating pitch based on the information of tones of speech and sex of speaker is introduced, too. At last, the modified parameters are imported to the system to get synthesized normal speech for completing the reconstruction of normal speech from Chinese whispers. The subjective test is done to evaluate the synthesized normal speech. It is shown that the method which is proposed in this thesis is feasible.