Implementing Thai text-to-speech synthesis for hand-held devices

The quality of Thai text-to-speech (TTS) have greatly improved over the past decade. To expand its utilization over platforms with different computational capacities, this paper described a real implementation of Thai TTS system for hand-held devices. The system consisted of two parts: the text analysis part and the speech synthesis part. To overcome an issue on the limited resources of the devices, a rule-based technique was chosen for the text analysis while the speech synthesis part was deployed in a model-based approach, specifically hidden Markov models, rather to the concatenative one. Parameters of the synthesis part were experimented with to reduce the computational time. The execution time and subjective mean opinion scores (MOS) have been measured. Experimental results show that computational requirements can be reduced by shorten the length of synthesis filter impulse response and the dimension of the feature vectors to some degrees without sacrificing the synthetic speech quality.