A statistical procedure for classifying word-initial voiceless obstruents is described. The data sel to which the analysis was applied consisted of monosyllabic words starting with a voiceless obstruent. Each word was repeated six times in the carrier phrase "I can say __, again" by each of ten speakers. Fast Fourier transforms (FFTs), using a 20-ms Hamming window, were calculated every 10 ms from the onset of the obstruent through the third cycle of the following vowel. Each FFT was treated as a random probability distribution from which the first four moments (mean, variance, skewness, and kurtosis) were computed. Moments were calculated from linear and Bark transformed spectra. Data were pooled across vowel contexts for speakers of a given gender and input to a discriminant analysis. Using the moments calculated from the linear spectra, 92% of the voiceless stops were classified correctly when dynamic aspects of the stop were included. Even more important, the model constructed from the males' data correctly classified about 94% of the voiceless stops produced by the female speakers. Classification of the voiceless fricatives when all places of articulation were included in the analysis did not exceed 80% correct when the moments from either the linear or Bark transformed scales were used. However, classification of only the voiceless sibilants was 98% correct when the moments from the Bark transformed spectra were used. As with the stops, the classification model held across gender.

A broad review of literature describing lingual function during speech shows that speaker samples per study are typically small (N<3 in more than 80% of all cases), and that speech samples, and representational and analysis conventions are highly variable. Similar conclusions can be drawn for other articulators. Thus it is fair to argue that there is still not available any valid, statistically-defensible sense of normal speech motor behavior, against which disordered articulatory behavior can be compared. Accordingly, a large-sample, 50-speaker x-ray microbeam speech database will be developed at the University of Wisconsin, incorporating point-parametrized representations of lingual, labial, mandibular, and velar movements in association with the resulting acoustic sound pressure wave, for a rich set of utterances and oral motor tasks, and lengthy recording interval (circa 18 min/speaker). The database is intended to be uniform across speakers in task inventory and descriptive kinematic framework; sufficiently accurate and deep to withstand scrutiny of variance, within and across speakers, and perhaps most importantly, an open source available for unlimited inspection and use by other speech scientists. Descriptions of the proposed speech inventory, experimental protocol, speaker sample, and timetable for database development will be provided. [Work supported by NIH DC00820.]

No abstract

A signal processing technique is described for measuring the jitter, shimmer, and signal-to-noise ratio of sustained vowels. The measures are derived from the least mean square fit of a waveform model to the digitized speech waveform. The speech waveform is digitized at an 8.3 kHz sampling rate, and an interpolation technique is used to improve the temporal resolution of the model fit. The ability of these procedures to measure low levels of perturbation is evaluated both on synthetic speech waveforms and on the speech recorded from subjects with normal voice characteristics.