Diane Kewley-Port Bill Mills Jonathan Dalby
Commun. Disord. Technol., 205 S. Walnut, Bloomington, IN 47404
Speaker-independent speech recognition can be used in speech training applications to provide feedback on phonemic contrasts. Our application is drill for normal-hearing misarticulating children with target-substitution errors (e.g., /r/->/w/ or /s/->/(theta)/). Six word pairs incorporating eight common misarticulations in word initial, medial, and final positions were repeated twice for a total of 192 words. These phonemic contrasts are extremely challenging for most recognizers. Seventeen children in third to fifth grade repeated the word list once. Two other children and an adult repeated it seven times to provide speaker-dependent comparisons. The Scott SIR 20 recognizer was evaluated. Speaker-independent templates were made in round-robin fashion, leaving out one child each time for testing. The overall recognition rate for the 16 tests was 75.3% correct, with a range of 85.3% for the best (/s--(sh)/) phonemic contrast, to 65.5% for the worst (/d--(delta)/). Human recognition of a subset of the words was also shown to vary across phonemic contrasts. Analyses of the linguistic content of the word pairs and comparisons with the speaker-dependent recognition results indicate that speaker-independent recognition is accurate enough for some minimal phonemic contrasts to be used in speech drill. [Work supported by NIDCD.]