A multiresolution ABS/OLA sinusoidal model using wavelet transform

The sinusoidal model has been applied to a broad range of speech and audio coding, such as analysis/synthesis, time-and frequency-scale modification, fundamental frequency modification, speech enhancement, and co-channel separation. In this paper, we propose a multiresolution analysis-by-synthesis/overlap-add (ABS/OLA) sinusoidal model using a wavelet transform. In the proposed scheme, after decomposing an input speech signal into multiresolution subband signals using the wavelet transform, classical ABS/OLA sinusoidal models with different window lengths are applied to each subband signals respectively. It is shown that by applying a proper-sized analysis window, much more accurate sinusoidal components can be estimated. Experimental results have shown that the proposed multiresolution ABS/OLA sinusoidal model can achieve better performance than that of the classical ABS/OLA sinusoidal model in terms of the spectral characteristics, phase characteristics, and the quality of synthetic speech.