Speech analysis and synthesis with a refined adaptive sinusoidal representation

This paper explores common speech signal representations along with a brief description of their corresponding analysis–synthesis stages. The main focus is on adaptive sinusoidal representations where a refined model of speech is suggested. This model is referred to as Refined adaptive Sinusoidal Representation (R_aSR). Based on the performance of the recently suggested adaptive Sinusoidal Models of speech, significant refinements are proposed at both the analysis and adaptive stages. First, a quasi-harmonic representation of speech is used in the analysis stage in order to obtain an initial estimation of the instantaneous model parameters. Next, in the adaptive stage, an adaptive scheme combined with an iterative frequency correction mechanism is used to allow a robust estimation of model parameters (amplitudes, frequencies, and phases). Finally, the speech signal is reconstructed as a sum of its estimated time-varying instantaneous components after an interpolation scheme. Objective evaluation tests prove that the suggested R_aSR achieves high quality reconstruction when applied in modeling voiced speech signals compared to state-of-the-art models. Moreover, transparent perceived quality was attained using the R_aSR according to results obtained from listening evaluation tests.