High resolution frequency analysis of voices--Feature extraction of nasal consonants

This paper gives a high resolution spectral analysis of Japanese nasal syllables uttered by 50 male and female speakers, feature extraction and recognition experiments. The spectral analyzer consists of a band-pass filter bank whose resolution is 24 or 96 channels per octave and Q=30 or 1 4. Detailed observation of spectral patterns of the nasal syllables shows that the features which distinguish between labial and alveolar exist not only in the second formant transition but also in frequency interval between the second and higher formants. In the recognition experiments for 1,000 nasal syllables based on a new recognition method named augmented feature space method (AFSM), all syllables are recognized correctly excluding several misheard ones in the hearing test.