Automatic evaluation of synthesized speech
This paper presents the development of a Croatian text-to-speech (TTS) system capable of synthesizing speech from arbitrary text. For speech generation pitch synchronous overlap and add (PSOLA) and hidden Markov models (HMM) based methods were used. The generated Croatian speech was evaluated subjectively by the mean opinion score (MOS) of the 40 evaluators and objectively by the automatic speech recognition (ASR) system. In the paper we propose an evaluation approach which combines objective and subjective evaluation results.