Thresholds for vowel formant discrimination using a sentence classification task

Abstract

Accurate classification of vowels in sentences is challenging because American English has many acoustically similar vowels. Using a 2AFC paradigm, our previous research estimated thresholds for vowel formant discrimination in sentences that were two times smaller than the measured formant distance between close vowels. A new paradigm has been developed to estimate the ability to detect formant differences in a sentence classification task. A seven-token continuum of changes in either F1 or F2 was generated from natural productions of ‘‘bid’’ and ‘‘bed’’ using the synthesizer STRAIGHT. These tokens were spliced into a nine-word sentence at different positions that also contained two other test words, one each from pairs ‘‘cot/cut’’ and ‘‘hack/hawk.’’ Listeners were asked to identify the three words they heard in the sentence. Listeners also identified whether ‘‘bid’’ or ‘‘bed’’ was heard when only the isolated tokens were presented. Thresholds to detect a change from ‘‘bid’’ were obtained from psychometric functions fit to the data. Thresholds were similar for the sentence and word-only tasks. Overall, thresholds in both classification tasks were worse than those from the 2AFC tasks. Results will be discussed in terms of the relation between these discrimination thresholds, vowel identification, and vowel spaces. [Work supported by NIHDCD-02229.]