Improve music synthesis quality by particular harmonic spectrum interpolation based on sinusoidal model

Sinusoidal model is not only widely applied in speech processing, but also an effective method in the analysis and synthesis of music signals. The method of peak parameters tracking largely matches the spectrum of harmonic components. However, it ignores much music particular harmonic spectrum components, which affects the clarity and the auditory effect of the synthetic music. This paper proposes a particular harmonic spectrum interpolation (PHI) method to compensate the deficiency of peak tracking, we use cubic spline curve interpolation to insert information of the particular harmonic spectrum between peaks, then the peak spectral parameters and interpolation particular spectrum parameters are transferred to music signal synthesis generator. Experiments show that the LSD (log spectral distance) is reduced by 5.3%, as well as the MOS score is improved by 0.4 in average comparing with traditional peak tracking. Furthermore, the auditory perception of synthesis music is largely improved, quite closer to the original signal.