Epochs based compression of LP residual for source modeling in text-to-speech synthesis

The objective of this work is to reduce the data rate of LP residual for source modeling in Text-to-speech synthesis (TTS) using the knowledge of epochs present in the speech signal. Epochs here refers to glottal closure, glottal opening, onset of bursts and some high amplitude instants in fricatives. Epochs are identified using both zero frequency filter (ZFF) and Hilbert envelope (HE) of LP residual. By using epoch as anchoring point, residual samples around the epoch are considered for compression of LP residual termed as residual block. By varying the number of samples in residual block around the epochs, data rate of LP residual is reduced and it is experimentally studied for synthesis interms of both naturalness and perceptual distortion. Finally, the proposed method is compared with compression of LP residual using Discrete cosine transform (DCT). Experiments show that naturalness of synthesized speech is comparable with the compression using DCT.