Period information deviation on the segmental sinusoidal model

Abstract

Speech signal can be modeled by sinusoidal model. On the sinusoidal model, there are many kinds for representing the signal. One of model is Segmental Sinusoidal model. The segmental sinusoidal model is an approximation method based on sinusoidal model for speech signal, especially for periodic detection. The periodic signal can be decomposed by infinite sinusoidal signal with combination of amplitude, frequency and phase. After the signal is decomposed, parameter will be quantized. The proposed quantization method in this paper is sampling signal on big part between minimum and maximum part over observation block. Some parameters of speech signal are detected. The useful parameters are peaks and period between consecutive peaks. Period information obtained from this quantization tends to different than the original, In this paper, we show the experimental results that there are many differences between period information on encoder side with the decoder side. It caused by quantization error on period information and quantization error on the codebook design. Effect of differences is degradation of signal quality, especially on frequency signal accuracy. On this paper, deviation of the reconstructed signal from original signal will be evaluated. Deviation from the original signals means that some error occur on period quantization.