Improving Fluency in a Sesotho Text-to-Speech Hybrid System

Most of the present text-to-speech systems produce an acceptable quality speech output. Text-tospeech systems that are based on limited domain techniques produce speech that is close to human speech; however, they lack flexibility in that they cannot be used to synthesize words not in their own vocabulary. One approach of dealing with the flexibility problem is to use hybrid systems which combine limited domain systems and open vocabulary systems. This only solves part of the problem as discontinuities between words generated by different systems become apparent in the produced speech. In this paper, we improve the hybrid system by implementing techniques that can mask the discontinuities so that the output speech is more fluent. The proposed system was evaluated by carrying out subjective listening tests. In the tests, 20 listeners evaluated the quality of the speech output based on the MOS scoring system. The results showed an improvement on fluency with an overall score of 3.7 from 3.05.