ASA 126th Meeting Denver 1993 October 4-8

4aSP9. A diagnostic text-to-speech intelligibility test.

Ann K. Syrdal

AT&T Bell Labs., Rm. 3E-525, 101 Crawfords Corner Rd., Holmdel, NJ 07733

Beki A. Sciacca

AT&T Bell Labs., Naperville, IL 60566 and Northwestern Univ., Evanston, IL 60208

A wider sampling of speech segments and contexts is required for adequately evaluating text-to-speech output than is typically used in standard intelligibility tests of human speech. The test developed expanded upon both the Diagnostic Rhyme Test (DRT) [W. D. Voiers, Speech Tech. (Jan/Feb.), 30--39 (1983)] and a TTS intelligibility test developed by van Santen [J. P. H. van Santen, Comp. Speech Lang. 7, 49--100 (1993)]. The featural system used to construct the test ensured that the most likely perceptual confusions were potential test items. One- and two-feature contrasts were tested in word-initial, word-medial, and word-final positions, in stressed and unstressed syllables. Vowel, consonant, and consonant cluster contrasts plus insertions and deletions were tested. Test items were presented in semantically anomalous sentences (e.g., ``Happy mimes dance the book.'') and were controlled for familiarity and verb transitivity. Testing was automated and conducted interactively with individual subjects who received no feedback on their performance. Results indicate the test was a useful diagnostic tool as well as evaluation technique.