NSF Project IIS-0535168: Separating Speech from Speech Noise Annual Report 2007
Pierre L. Divenyi, DeLiang Wang · 2008
This year we continued our multisite collaboration aimed at improved understanding and modeling of how listeners handle noisy or corrupt speech. On the perception side, we have continued experiments on how listeners perceive distorted and incomplete speech tokens, including the use of our spondee/foil paradigm, in which listeners are played tokens derived from two-syllable compound words which are either relatively familiar spondees (e.g. “bankroll”), or pairs of monosyllable words less likely to be encountered together, but with the same initial consonant-vowel, and final vowel-consonant (e.g. “band hole”). Our manipulations consist of replacing parts of the original speech with short acoustic excerpts that provide reduced information, for instance by replacing a 200 ms chunk with 200 ms of white noise. In the spondee/foil paradigm, we generally replace a section of the speech centered on the consonant cluster between the two syllables, which tends to mask the distinction between spondee and foil. Increasing the temporal width of the replacement, and/or reducing the amount of information from the original speech preserved in the replacement, will lead to