The Application of Speech Synthesis Technology Based on Deep Neural Network in Intelligent Broadcasting

Abstract

To improve the sound quality of speech synthesis technology in intelligent broadcasting, a deep neural network-based method is proposed. It also proved the effectiveness of the DNN discrimination s/u/v and completed the conversion of the HMM synthesis spectrum parameter to original speech. Further, the scheme for transforming the parameters obtained from the temporary decomposition (TD) algorithm, DNN trains the event vectors obtained from TD decomposition, establishes the transformation model, and recombines with the untransformed event function. Experiments proved that the conversion effect of 16 dimensional parameters is not very ideal in subjective evaluation due to the fact that too few dimensions lead to insufficient spectral details, and the distortion in the process of further synthesis; the parameter conversion of 48 dimensions is slightly better than 16 dimensions, mainly due to more spectral details, but on the other hand, the influence of codebook mapping also affects the sound instability to some extent. It proves that the intelligent voice broadcast system completely solves these problems, which not only reduces construction costs, but also improves service efficiency.