Third International Conference on Spoken Language Processing (ICSLP 94)
In this paper, we describe a method to generate non-entry words from entries of the natural speech database which an automatic public announcing system is possessed of. Thereby, it becomes unnecessary to record new voices by a narrator. Non-entry words are generated by means of the waveform editing, that is, by the method of segmental speech sound concatenation. In case that there is no need to change the pitch pattern at editing, quality of the generated words is maintained to the level of natural speech sound. In case that the pitch pattern must be changed at editing, the zero-phased pitch waveform superposing method is used for pitch modification. In order to extract raw pitch waveforms, various windows including the Hanning and the Blackman-Harris whose length are proportional to the pitch period are tested. And, following results are obtained: (1)The Hanning window whose length is twice the pitch period is slightly superior to the Blackman-Harris windows. (2)Quality degradation of the generated words is a little bit.
Bibliographic reference. Arai, Yasuhiko / Minowa, Toshimitsu / Yoshida, Hiroko / Nishimura, Hirofmi / Kamata, Hiroyvki / Honda, Takashi (1994): "Generation of non-entry words from entries of the natural speech database", In ICSLP-1994, 1291-1294.