Statistical modification based post-filtering technique for HMM-based speech synthesis

The speech generated from hidden Markov model (HMM)-based speech synthesis systems (HTS) is suffered from over-smoothing problem which is due to statistical modeling. This paper will focus on post-filtering technique based on statistical modification for the generated speech parameters. The marginal statistics of parameters´ trajectory, such as mean, variance, skewness and kurtosis are adjusted according to the values generated from the HTS system. This technique is compared with global variance (GV)-based speech generation algorithm. The listening test showed that the post-filtering technique considering the mean and variance could generate almost equal result with GV model. When further considering the modification of skewness and kurtosis, the quality of generated speech has been improved.