Post-processing of speech signal for prosody modification and improvement

Abstract

The basic task of a text-to-speech (TTS) synthesis system is to obtain the correct synthetic speech signal with the help of machines corresponding to the given input text. However, the main difficulty with the TTS system is the problem of appropriate prosody in the resultant speech signal. In this thesis, we used the methods based on the pitch synchronous overlap-add (PSOLA) technique, i.e., time-domain PSOLA (TD-PSOLA) and linear prediction PSOLA (LP-PSOLA), which tries to use the combination of different pitch-scale and time-scale combination to match the synthesized speech to the natural speech. To implement the PSOLA techniques, different pitch detection algorithms are employed in order to obtain the pitch marks and pitch contour. Pitch marking is essential task to obtain the required time-scale and pitch-scale modifications. Pitch detection algorithms based on autocorrelation function (ACF), normalized cross-correlation function (NCCF) and zero frequency resonator (ZER) are employed in this thesis. Firstly, we applied the PSOLA methods to the unit selection synthesis (USS) and Hidden Markov model-based TTS (HTS) based synthesized speech for which we were having the prior knowledge of natural speech corresponding to the synthesized speech. Later, we performed the method on the Blizzard Challenge-2012 speech corpus for which we were not having the database of corresponding natural signal. PSOLA method is also applied only on the natural speech for time-scale and pitch-scale modifications. Time-scale modification of natural speech have many real world applications speech, a series of tests are then performed to determine the effectiveness of the PSOLA methods.