The Research Of Music Synthesis Based On Sinusoidal Model

Sinusoidal model is widely used in the synthesis of speech signals.In this paper,we study further the music synthesis based on sinusoidal model.The traditional method of peak tracking of sinusoidal model has certainly theoretical basis,but the quality of music reconstruction is not high.Though the method of peak tracking largely fits harmonic spectrum,it lost much non-harmonic spectrum which affect the music signal a lot.To avoid the shortage,a method is proposed in this paper by using particular harmonic interpolation(PHI)which interpolates particular points according to the relationship between adjacent peak parameters.Then the cubic spline curve is used to fit improved parameters spectrum,this method enriches the spectral structure which influences instrument timbre,and makes the synthesis quality greatly improved.Finally,subjective and objective evaluations are operated in order to verify of the effectiveness of the method.Compared with the traditional method,the average value of LSD(Log Spectrum Distance)is decreased by 5.3%,and the value of MOS score is average increased by about 0.365.Secondly,the paper takes the audio storage and low bit rate transmission into consideration,then we take use of the characteristics of wavelet decomposition of the music signal which make the energy concentrated in the low frequency coefficient.We gain the effective coefficient by soft threshold to compress data,and the compressed signal can reconstruct the music signal combined with PHI based on the sinusoidal model.The tests are done to prove that it can keep the reconstructed music with the high quality with compressed signal data.