First International Conference on Spoken Language Processing (ICSLP 90)
Listeners hearing distorted speech may be able to "correct" it if they have enough information about the nature of the distortion. We present evidence that this ability to correct a distorted speech signal is bought at a slight cost, namely, hyper-correction. We presented American English listeners with brief (85 msec) samples of 11 vowels that had been 1 kHz low-passed filtered. When presented in a way that gave them no chance to learn the filter characteristics they made 33.3% correct identifications. Many of the errors involved front vowels (those having a high Formant 2) being confused with back vowels (which often have F2 and F1 fused at 1 kHz or less). When the samples were preceded by a redundant precursor sentence filtered in the same way as the vowel samples, listeners raised their correct identifications to 50.1% but there was an increase in "hyper-correction" errors, i.e., where back vowels were identified as front vowels. This may provide clues as to how listeners "correct" distorted signals.
Bibliographic reference. Ohala, John J. / Shriberg, Elizabeth E. (1990): "Hypercorrection in speech perception", In ICSLP-1990, 405-408.