Cochlea-scaled entropy predicts intelligibility of Mandarin Chinese sentences

Abstract

Cochlea-scaled spectral entropy (CSE; Stilp & Kluender, PNAS, 107(27):12387-12392 [2010]) is a measure of information-bearing change in complex acoustic signals such as speech. CSE robustly predicts English sentence intelligibility even amidst temporal distortion and widely different speaking rates (Stilp, Kiefte, Alexander, & Kluender, JASA, 128(4):2112-2126). However, CSE does not explicitly capture changes in fundamental frequency (f0) in any way distinct from that for other aspects of spectral shape (e.g., formant patterns, slope). This property of CSE could limit its ability to predict intelligibility of tone languages that use f0 for phonetic and morphological distinctions. The present study assesses the predictive power of CSE for Mandarin Chinese sentence intelligibility. Twenty-five native-Mandarin listeners transcribed Mandarin sentences in which consonant-length (80-ms) and vowel-length (112-ms) segments with either high or low CSE were replaced with speech-shaped noise. CSE reliably predicted listener performance; as greater amounts of CSE were replaced by noise, performance worsened. Results encourage information-theoretic approaches to speech perception, as change and not physical acoustic measures best predict sentence intelligibility across tonal and nontonal languages. [Work supported by NIDCD.]