LSF and Phase Feature Combination for Join Cost Estimation in a TTS System

In recent text-to-speech (TTS) synthesis systems, speech is generated by concatenating units that are selected from a large database. Selection of units can be based on two cost measures: join and target. Many methods have been developed for join cost estimation, but they all suffer from neglecting the phase feature. Our new approach for join cost estimation is based on combining the residual phase feature and LSF feature where Discrete All Pole modeling(DAP) has been used for phase and LSF feature extraction. Results of different experimentations show that this method has better performance than conventional methods like MFCC unit selection.