Nasal detection module for a knowledge-based speech recognition system

The Lexical Access From Features (LAFF) project tries to model the representation and perception of speech by human listeners. The derivation of such a representation involves first finding certain acoustic landmarks. Based on the landmarks and the acoustic cues surrounding the landmarks, distinctive features of the speech segments may be deciphered. The present study concentrates on the nasality module that attempts to detect the presence of an underlying nasal consonant, which is almost always adjacent to a vowel. For an underlying nasal in English, the features [+voiced, +sonorant, +consonant, +nasal, -continuant] are specified. The features are then mapped into measurable acoustic properties. Normally, cues from three regions in the sound indicate the presence of a nasal consonant: (1) abrupt spectral change from the vowel to the nasal murmur, (2) vowel nasalization, and (3) nasal murmur. These cues are quantified by acoustic parameters whose values are combined to indicate the presence of a nasal. The nasality module that has been developed is a sonorant landmark detector that greatly reduces false landmark detection and distinguishes nasals from laterals by incorporating additional nasal manner cues. The module also addresses cases where one or more of the three nasal cues is absent.