On methods for perfect reconstruction WI speech coding with preprocessing

Abstract

The waveform interpolation (WI) speech coding algorithm has been shown to be an efficient method to describe the evolution of periodic voiced components in the speech signal. However, the conventional WI coding does not provide perfect reconstruction property, i.e. the decoded signal does not converge to the original signal with decreasing quantization error. Therefore errors in the coding model cannot be fixed by quantization. In this paper we discuss about characteristics of the WI coding model and about modifications to the model which enable the perfect reconstruction property. The new requirements and features are examined and discussed in detail. While the perfect reconstruction property brings many benefits it also causes new demands to the operation of the coder. Particularly high requirements are set to the exactness of the pitch estimate; inaccuracies hamper rapidly the possibilities to quantize the parameters efficiently. To overcome this we introduce a preprocessing method which slightly modifies the pitch structure of the residual signal before waveform extraction. The modifications to the signal are minor and therefore the quality of the preprocessed signal is very close to that of the input speech. In the proposed method the perfect reconstruction property is maintained in relation to the preprocessed signal.