Vocal microtremor in normophonic and mildly dysphonic speakers

1. Vocal microtremor in normophonic and mildly dysphonic speakers Jean Schoentgen Université Libre Bruxelles Brussels - Belgium

2. Vocal microtremor (definition) • Modulation of the phonatory frequency • Distinct from pathological vocal tremor • 1 - 15 Hz (Titze, 1995, Sataloff, 1997) • Two features : modulation level and modulation frequency

3. Vocal microtremor (examples)

4. Motivation ? • Tremor data are scarce • Vocal jitter & microtremor are base-line phenomena • Measurement of vocal tremor frequency via the cycle length time series • Test predictions of a simulation model of jitter and tremor (Schoentgen, 2001)

5. Experiment I : Objectives ? • Recording data (tremor level & frequency) • Differences between vowel timbres ? • Differences between male & female speakers ? • Differences between normophonic & mildly dysphonic speakers ?

6. Corpora • Sustained vowels [a], [i], [u] • 22 males, 16 females (normophonic) • 16 males, 28 females (dysphonic) • Voice type : monocycle periodic • Register : modal • No register or type breaks, or voice arrests • No cycle length outliers • No excessive additive noise or jitter • No pathological vocal tremor

7. Method (tremor frequency) • Estimation of the average cycle length • Upsampling (160 kHz) and low-pass filtering of the speech signal • Extraction of the vocal cycle length time series via peak picking • Removal of frequency drift or glissando • Calculation of the magnitude spectrum of the time series • Search for the statistically significant spectral peaks • Tremor frequency = weighted average of spectral peak positions

8. Examples of spectra

9. Method (tremor level) • Upsampling (160kHz) and low-pass filtering of the speech signal • Extraction of the vocal cycle length time series via peak picking • Removal of frequency drift or glissando • Smoothing of the time series to decrease jitter • Tremor level -> standard deviation of the smoothed cycle length perturbations (divided by average cycle length)

10. Results (1) • No statistically significant differences for the modulation frequency (Hz) and modulation level (%) between : • Male & female speakers • Normophonic & mildly dysphonic speakers • Vowel timbres

11. Results (2)

12. Results (3) • Dissimilarities between modulation data reported by different studies are due to different cutoff frequencies below which spectral peaks are considered not to contribute to vocal microtremor

13. Experiment II : Objective • Compare the size of vocal cycle length perturbations owing to jitter and frequency tremor

14. Corpus & Method • 22 male and 16 female speakers sustained [a], [i] and [u]. • Upsampling (160kHz) and low-pass filtering of the speech signal • Extraction of the vocal cycle length time series via signal zero-crossings • Removal of frequency drift or glissando

15. Linear auto-regressive analysis of the cycle length time series (e.g. Schoentgen, 1995) • Present perturbation = weighted sum of past perturbations + de-correlated noise • De-correlated noise -> vocal jitter • Weighted sum -> vocal tremor (by default) • Calculate sample standard deviation for each (& divide by average cycle length)

16. Example

17. Results (1) : [a](inter-quartile ranges)

18. Results (2) • Vocal jitter (%) < vocal tremor (%) (statistically significant) • Moderate significant correlation between vocal jitter & tremor (in %) • No significant tremor differences between vowel timbres • No significant tremor differences between speaker genders

19. Conclusion • Vocal frequency (micro)tremor data can be obtained via the cycle length time series • This may be generalized to pathological tremor data, but an additional stage may be required which is the re-sampling of the cycle length time series at equal intervals