Third European Conference on Speech Communication and Technology

Berlin, Germany
September 22-25, 1993


Auto-Regressive Linear Models of Jitter

Jean Schoentgen (1,2), Raoul de Guchteneere (1,3)

(1) Institute of Phonetics, CP110, Université Libre de Bruxelles Brussels, Belgium
(2) National Fund for Scientific Research, Belgium
(3) Grant U.L.B.

Jitter is defined as the fluctuations of the lengths of the glottis cycles during sustained phonation. Generally speaking, the amount of jitter is estimated by measuring successive glottal cycle durations and by computing a dispersion measure over an analysis interval of typically fifty cycles. A well-known descriptor of jitter is the so-called Period Perturbation Quotient (PPQ). The problem with this approach is that measures of dispersion do not adequately represent the chronological aspect of period sequences. In other words, the fact that fluctuations of consecutive period may be correlated is not taken into account. Therefore, we decided to study jitter by statistical time-series analysis methods, which are able to represent the interdependences of the fluctuations of neighbouring cycles. Accordingly, we developed an algorithm that measured cycle durations with great precision and used a stochastic linear model to represent cycle duration sequences. The model was an autoregressive linear model. The order of the model was determined on the basis of the partial correlation coefficients and it was confirmed by statistical tests of the whiteness of the residue. We fitted the model to cycle durations of the sustained vowels [a], [i] and [u] produced at a comfortable pitch and loudness level by 40 healthy speakers and 15 dysphonic speakers. Then, we calculated the PPQs of the whitened and unprocessed cycle fluctuations sequences. Results showed that the fluctuations of adjacent periods of a large majority of our vowel signals were correlated. When the fluctuations of adjacent cycles were strongly correlated, especially negatively, the PPQ of the residue was smaller than the PPQ of the raw sequence. By contrast, PPQ values of the whitened and unprocessed sequences were comparable when the correlation between adjacent cycles was weak. Also, AR model order and PPQ values appeared to be different for the three vowel qualities, for dysphonic and healthy speakers, and for male and female speakers.

Keywords: Jitter, AR-models

Full Paper

Bibliographic reference.  Schoentgen, Jean / Guchteneere, Raoul de (1993): "Auto-regressive linear models of jitter", In EUROSPEECH'93, 2033-2036.