Optimizing HMM Speech Synthesis for Low-Resource Devices
Bálint Tóth, Géza Németh · Journal of Advanced Computational Intelligence and Intelligent Informatics · 2012
Speech synthesis can be an importantmodality in Cognitive Infocommunications (CogInfoCom). Speech output is beneficial when the visual output of a system is blocked or is difficult to reach. Extra information can be added to output by applying different voice characteristics and emotional speech. CogInfo-Com systems can use low-resource devices in many cases. This paper describes the application of Hidden Markov Model (HMM) based speech synthesis to such systems. Several optimization steps, e.g., changing HMM parameters, applying performance-specific programming methods, are analyzed on three different smartphones in terms of speed, footprint size, and subjective speech quality. The goal is to approach realtime functionality while keeping the speech quality as high as possible. Successful optimization steps and resource-dependent optimal settings are introduced.