Research On Reconstruction Of Chinese Normal Speech From Whispered Speech Based On RBF Neural Network

Whispered speech is different from the normal speech in the pronunciation pattern. The previous researches on whispered speech are mainly about voice basic study or the needs of medical work. As the mobile communication products are used widely at present, whispers are used to avoid influencing others or being overheard. Because the phonation energy of whispers is low and its intelligibility is bad, it's difficult for people to communicate with each other clearly. The problem can be resolved if the mobile communication products have the ability of converting the whispered speech into the normal speech. Besides, this technology can also help aphonic individuals and the patients with pathological voice to communicate normally.The research of this paper is mainly about reconstruction of Chinese normal speech from whispered speech based on RBF neural network. Main work includes:Based on analyzing acoustic characteristics of whispered speech, study about the presented conversion method such as the linear prediction coding (LPC) and the homomorphic signal processing (HSP).But the intelligibility and quality of the converted speech are bad.As the characteristic parameters of the whispered speech are different from the normal speech, this paper this paper establishes the relationship of spectral mapping between whispered and normal speech by using RBF neural network, which will modify the Linear spectral pair(LSP) parameters.Firstly, detect the endpoint and the initial/final of the whispers, pretreat the LSP parameters of the whispered and the normal speech; secondly, generate the nonlinear mapping of spectral envelope from whispered speech to the normal speech by RBF neural network, where the LSP characteristic parameters of the whispered speech are the inputs of the RBF neural network, and the LSP characteristic parameters of normal speech are targets of the output; thirdly, transform the spectral of the whispered speech by using the trained RBF neural network, add the average pitch, and convert the speech; finally, connect the initial consonant of the whispers, export the normal speech.Both subjective and objective ones are conducted on the converted speech quality by the methods. Simulation results show that the score of the mean opinion score (MOS) is 3.5; the distorted distance of bark spectrum is decreased. Both intelligibility and quality of the converted speech are satisfied.At last, this paper raises the shortcomings of this method and the problems that haven't been solved, and gives the direction of further study and improving.