ASA 126th Meeting Denver 1993 October 4-8

1pSP8. The role of importance sampling in the establishment of normal ranges for speech characteristics.

Jon H. Lemke Hani Samawi

Dept. of Preventive Med. and Environ. Health, 2800 Steindler Bldg., Univ. of Iowa, Iowa City, IA 52245

Most measurements of voice characteristics have inherently skewed distributions that have never been characterized for individuals without speech disorders, that is, individuals with ``normal'' voices. Any diagnostic value these measures might have requires that one studies the characteristics of these ``normal'' voice distributions, in particular the quantiles necessary to establish the normal range. The focus will be on perturbations measures, such as jitter and shimmer. To solve this problem a method of importance sampling has been developed for estimating quantiles of skewed distributions. Importance sampling is a modified bootstrap procedure with exponential tilting, that is, a resampling method where the probabilities of the original observations appearing in the new samples are no longer equally weighted but are weighted by their closeness as an order statistic to the quantile being estimated. To estimate the normal limits of jitter and shimmer without resampling procedures will require at least 800 subjects. However, to estimate the normal limits with the same efficiency using this modified importance sampling technique will require approximately 80 subjects. The assessment of voice is very expensive and time consuming. There will never be 800 normal voice assessments conducted under identical conditions by anyone. How modified importance sampling can be used to establish normal limits for jitter and shimmer will be demonstrated with data collected at the Denver Center for the Performing Arts, Denver, Colorado, and the Wendell Johnson Speech and Hearing Center, University of Iowa, Iowa City, Iowa.