Reducing Computational and Memory Cost for HMM-Based Embedded TTS System

Abstract

In this paper, we present several methods to reduce the computational and memory cost to embed HMM-based TTS system. We firstly decrease the number of HMMs by applying decision tree based context clustering technique. Secondly propose address-based model compression technique to compress the model size without degradation in synthesis speech quality. Thirdly reduce the feature vector size to decrease computational and memory resources. Finally, fixed-point implementation is taken to fit the TTS system requirements to embedded devices’ resource. Experimental results show that the system size can be compressed to 3.61MB from 293MB, memory and computational cost are low enough for real-time embedded application. Subjective evaluation shows that the synthesis speech quality is fairly good.