Some Properties of Japanese Sounds through Perceptual Experiments and Spectral Analysis.

Abstract

We have been studying on automatic speech recognition and our experience taught ourselves that we had better investigate "What features are important in phoneme recognition?", "Why is phoneme recognition difficult?" and "To which extent is it difficult?" for developing a successful speech recognition system. In this paper, we would try to answer to such questions through perceptual experiments and statistical analyses of spectra. First, we found the following properties of voiced consonants from perceptual experiments. (1) The change of speech power (energy) along with time-axis is not the cue of perception of voiced consonants. (2) More phonemic information of voiced consonant is contained in glide than in a central part of consonant. (3) Even if the glide is taken out of a VCV utterance, the intelligibility is beyond 60%~70%. Second, we obtained the following results from statistical analyses of static features of voiced consonants. (4) The spectral difference between any two vowels is larger than that between two voiced consonants. (5) The spectral difference between two voiced consonants having the same manner of articulation is particularly small. (6) /N/, /y/, /w/, /m/, /n/, and /g/ are more influenced by the speaker-factor than by adjacent phonemes (or context). Third, we tried three-dimensional representation of consonants based on the dynamic features of spectra and investigated the relation among consonants in that space. Finally, we investigated the spectral change of consonants, caused by the factor of speaker and vowel. The results are : (7) /w/ and /z/ have the property of unvoiced consonants as compared with other voiced consonants. (8) The spectra of vowels and unvoiced consonants are not subject to the influence of the speaker-factor so much as voiced consonants. (9) The spectra of /u/, /i/ and consonants followed /i/ and /u/ are subject to the influence of the speaker-factor more than others. (10) However, there is no significant difference of speaker-factor among consonants for a long time interval.