Prosodic Stress, Information, and Intelligibility of Speech in Noise

摘要

Prosodic stress increases the salience of stressed syllables. The project investigated whether this property of speech is used by listeners for the understanding of spoken sentences presented in noise. Stressed syllables in the 720-sentence IEEE corpus were marked and envelope contours were generated to increase or decrease the level of speech-spectrum noise in synchrony with the occurrence of stressed syllables. Data from ten normal-hearing young listeners indicate that signal-to-noise ratio (SNR) of stressed syllables alone is a good predictor of the intelligibility of the whole sentence. A computational model was also developed by decomposing speech into eight low- rate basis functions inspired by articulatory gestures. This model was applied to test the hypothesis that the slowly varying basis functions may reflect the way listeners recover low-SNR or otherwise distorted speech segments between higher-SNR segments that stressed syllables represent. Initial results of the model suggest that a surprising degree of articulatory information is transmitted across periods during which acoustic information has been suppressed.