A novel pitch-lag search method using adaptive weighting and median filtering

Abstract

This paper presents a novel method to estimate the pitch-lag in a speech codec. The pitch-lag is related to the fundamental frequency of the speech signal and an accurate estimation of this parameter is important for the subjective quality of the synthesised speech. A common problem in speech codecs is that the estimation of the pitch-lag often produces a multiple or a sub-multiple of the true pitch value. When these incorrect pitch-lag values are used in speech synthesis the subjective quality of the speech is degraded. This paper presents an improved method where the estimation of the pitch-lag parameter is biased towards the pitch-lag values of the previous speech segments resulting in a consistent set of consecutive pitch-lag values and a high quality reconstructed signal. The classification of speech into voiced and unvoiced parts is used when tracking the pitch-lag values and adapting the pitch track centered weighting function.