Trend tying in the segmental-feature HMM

Abstract

We present a reduction method for the number of parameters in a segmental-feature HMM (SFHMM). If the SFHMM shows better results than the CHMM, the number of parameters is greater than that of the CHMM. Therefore, there is a need for a new approach that reduces the number of parameters. In general, the trajectory can be separated by the trend and location. Since the trend means the variation of segmental features and occupies a large portion of the SFHMM, if the trend is shared, the number of parameters of the SFHMM may be decreased. The proposed method shares the trend part of trajectories by quantization. The experiments are performed on the TIMIT corpus to examine the effectiveness of the trend tying. The experimental results show that its performance is the almost same as that of previous studies. To obtain better results with a small amount of parameters, the various conditions for the trajectory components must be considered.