Automatic detection of landmark for nasal consonants from speech waveform

Abstract

A knowledge-based approach towards automatically detecting nasal landmarks (/m/, /n/, and /ng/) from speech waveform is developed. The acoustic characteristics Fn1 locus calculated on each frame of speech waveform as the mass center of spectrum amplitude in the vicinity of the lowest spectral prominence between 150-1000Hz, and A23 locus calculated on the same speech frame as a band energy between 1000-3000Hz were incorporated together to construct the nasal landmark detector, which alarms at the instants of closure and release of nasal murmur. Experiment observations on the acoustic characteristics of Fn1 and A23 and the nasal consonant landmark detection results on the VCV database are also presented.