A method to choose an appropriate concatenating position for automatically generated synthesis unites

To make synthesized speech sounds more natural, one would prefer using a larger size of speech data base. However, when we adopt manual segmentation method in making a large size data base, the procedure would be very time-consuming and the constructed data base would be in consistent. Consequently, automatic segmentation using speech recognition system has been used for making a large size of speech synthesis data base. However, when the automatic segmentation were applied to make synthesis speech, significant concatenation distortion happens due to phoneme boundary error. The purpose of our study is to choose an appropriate concatenating position of automatically generated synthesis units which may have errors on their boundaries. We have performed MOS(Mean Opinion Score) tests and analyzed the shape of spectrograms to evaluate our proposed algorithm. Our test results have shown that the quality of concatenated speeches with proposed algorithm is superior to that of concatenated ones generated by automatic segmentation only.