Cochlea‐scaled entropy predicts speech intelligibility when intensity is scaled as linear amplitude or as loudness, but not when scaled logarithmically.

Abstract

Stilp and Kluender [Proc. Natl. Acad. Sci. 107, 12387–12392 (2010)] reported sentence intelligibility to be well‐predicted by measures of sensory change over time [cochlea‐scaled spectral entropy (CSE)] and not consonants, vowels, or duration of signal replaced. As greater amounts of CSE were replaced by noise, intelligibility worsened. They calculated CSE as Euclidean distances between spectra defined by linear amplitude as a function of equal rectangular bandwidth (ERB). The present experiments compare measures of speech intelligibility as a function of CSE when amplitude is measured on a linear scale as before and when log‐scaled (dB) or loudness‐scaled (sone). Listeners typed all words they understood after hearing sentences with 80‐ and 112‐ms segments replaced by speech‐shaped noise. Log‐scaled (dB) CSE was a poor predictor of performance. Both linear‐scaled and loudness‐scaled measures of CSE predict intelligibility very well, replicating Stilp and Kluender. Linear and loudness scales will be discussed in the context of upward spread of masking and two‐tone suppression as they relate to speech signals. These perceptual findings suggest that conventional use of spectrograms (linear frequency × dB intensity) could systematically misinform researchers concerning properties of speech signals that are most useful to listeners. [Work supported by NIDCD.]