Acoustic cues for the classification of re

Irregular phonation serves an important communicative function. It can be a cue to linguistic contrasts, and often serves as a marker for word and utterance boundaries. Automatic methods for classification and detection of regions of irregular phonation can be used to improve analyses of occurrences of irregular phonation and support technologies such as speech recognition and synthesis. This study proposes a set of acoustic cues from both the temporal and frequency domains — fundamental frequency, normalized RMS amplitude, smoothedenergy-difference amplitude and shift-difference amplitude— for separation of regions of regular and irregular phonation. Tokens from the TIMIT database are classified using support vector machines, trained on 114 different speakers and tested with 37 different speakers. Both genders are well represented in the data set and the tokens occur in various contexts within the utterance. In the test set, 292 of 320 irregular tokens (recognition rate of 91.25%), and 4105 of 4320 regular tokens (recognition rate of 95.02%) are correctly identified. [Work supported by NIH/NIDCD # DC02978.]