Management of Truncated Data in Speech Transmission Evaluation for Pupils in Classrooms

Abstract

Speech intelligibility is a subjective performance index defined as the percentage of a message understood correctly. Often the results of speech intelligibility tests would suggest that conditions are acceptable, with Intelligibility Score (IS) of the order of 90% or more, while speech transmission performance may not be satisfactory. Subjective ratings of the Listening Easiness Score (LES), based on a discrete questionnaire, provide an alternative approach. A total of 239 primary school pupils, aged 7 to 11, evenly distributed among the grades, participated in the survey. The objective indicator Speech Transmission Index (STI) was also measured for each test setting in seven different positions in the laboratory classroom used for the test. Both IS and LES are inherently bounded, and their data distributions exhibit a significant accumulation of scores in the upper and lower parts. The resulting truncation problem has been addressed with a method based on the normal probability plot, enabling identification of mathematical models relating IS and LES to STI, as well as the estimation of related uncertainties. IS and LES exhibit substantially similar metrological capabilities, as, for both, model relative uncertainty does not exceed 4% and uncertainties in prediction of new observations are about twice as large.