Toward an automatic system for distinctive feature extraction from speech
Andrew Wilson Howitt · The Journal of the Acoustical Society of America · 1990
A novel representation of acoustic information leading to the extraction of distinctive features is proposed. This representation is hierarchical and nonsegmental, and it is constructed in a three-stage process. First, frame-based parameters are calculated, which represent the spectral characteristics of the signal, with attention to the auditory representation of these characteristics. Second, events or discontinuities in the acoustic signal are detected (with the aid of the parameters from the first stage), which represent information about certain manner features. Third, information about laryngeal and place features is extracted via processing in regions based on the events from the second stage. Examples of the process of feature extraction by this method will be presented and comparison with other methods will be discussed. The proposed representation should be useful both as a front end for speech recognizers and as a research tool for the study of models of lexical access by human listeners.