ASA 130th Meeting - St. Louis, MO - 1995 Nov 27 .. Dec 01

5aSC4. Intelligibility of normal speakers: Error analysis.

Amy T. Neel

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

Errors made by listeners in sentence transcription were analyzed to determine their contributions to intelligibility. One hundred Harvard sentences produced by ten males and ten females were transcribed by ten listeners per talker. Two measures of intelligibility were obtained: a keyword score in which a sentence was correct if, and only if, all five keywords were correct; and a total error count. Analysis of error types revealed that typing/spelling errors accounted for a third of total errors, and phonetic errors (consonant and vowel errors) accounted for another third. The remainder were semantic errors, added or deleted words, or unclassifiable. Further analysis of consonant errors did not reveal any particular type of consonant to be more susceptible to error than others. Male talkers had significantly worse keyword scores than females but did not have significantly greater total error counts indicating more errors on function words for females. The difference between high and low intelligibility speakers (by total error count) was accounted for by increased typing/spelling and consonant errors. Acoustic analysis of incorrectly transcribed words revealed that phonetic errors appeared to originate in the mouths of speakers while errors like word substitutions arise in the ears (or brains) of listeners.