ASA 127th Meeting M.I.T. 1994 June 6-10

5pSP6. Using a multi-talker database to identify sentence- and talker-dependent correlates of speech intelligibility: Preliminary results.

Ann R. Bradlow

David B. Pisoni

Speech Res. Lab., Dept. of Psychol., Indiana Univ., Bloomington, IN 47405

Intelligibility tests of a database of 2000 recorded sentences (100 sentences produced by 20 talkers), show considerable variation in intelligibility across both sentences and talkers. This paper presents preliminary results of a study designed to identify the lexical and talker characteristics that correlate with speech intelligibility scores. To identify sentence-specific correlates of intelligibility, lexical characteristics of the words that comprise the most versus the least intelligible sentences were compared. Results show that high-intelligibility sentences have a higher proportion of function to content words, and consequently, a higher mean word frequency than low-intelligibility sentences. Furthermore, lexical characteristics of the most common word-level substitution errors indicate the importance of word frequency in overall sentence intelligibility. To identify talker-specific correlates of intelligibility, acoustic measurements of sentences that show the most inter-speaker variability in intelligibility were compared. For example, durational characteristics of the most common segment-level errors indicated that the relative timing between segments is an important aspect of talker intelligibility. Future directions and general implications of this database project will be discussed. [Work supported by NIH Training Grant No. T32 DC-00012 and NIH Research Grant No. R01 DC-00111.]