In a phoneme-based speaker-adaptive automatic recognition system for continuous English speech, a segmentation algorithm for nasals uses automatically derived thresholds on spectral energy measures. A Gaussian classifier using formant information, with duration, two compound measures, and a 'spectral contrast' measure, is applied to the hypothesised nasal segments, which are classified phonemically. Tests were carried out over the training data and over a second reading for each of two speakers. Results were comparable with earlier work as regards nasal/non-nasal hit rate, and better as regards the ratio of imposters to nasals. This method also goes further and achieves some success at the phonemic level.
Bibliographic reference. Williams, B. / Hiller, S. M. / McInnes, Fergus R. / Dalby, Jonathan (1989): "A knowledge-based nasal classifier for use in continuous speech recognition", In EUROSPEECH-1989, 2252-2255.