Decomposition of speech signals into a periodic and non-periodic part based on sinusoidal models

This paper will focus on the speech signals decomposition problem based on three sinusoidal models which can be viewed as three versions of a basic model: the Harmonic plus Noise Model, HNM. For the first model, noted as HNM1, the harmonic amplitudes are retained constants within the analysis frame; the phase is assumed to be linear. The second model, HNM2, makes use of linearly varying complex amplitudes for the harmonic part whereas, the third model, HNM3, makes use of a third order polynomial with real coefficients for the harmonic amplitudes and also linear phase as the first model. To obtain the non-periodic part (or noise part) the harmonic part is simply subtracting from the original speech signal in the time domain.