Intelligibility predictors and neural representation of speech

Intelligibility predictors tell us a great deal about human speech perception, in particular which acoustic factors strongly effect human behavior, and which do not. A particular intelligibility predictor, the Articulation Index (AI), is interesting because it models human behavior in noise, and its form has implications about representation of speech in the brain. Specifically, the Articulation Index implies that a listener pre-consciously estimates the masking noise distribution and uses it to classify time/frequency samples as speech or non-speech. We classify consonants using representations of speech and noise which are consistent with this hypothesis, and determine whether their error rate and error patterns are more or less consistent with human behavior than representations typical of automatic speech recognition systems. The new representations resulted in error patterns more similar to humans in cases where the testing and training data sets do not have the same masking noise spectrum.

Figure optionsDownload as PowerPoint slideResearch highlights
► A representation of speech is inferred from human behavior.
► That and other representations are used to recognize consonant-vowel syllables.
► Error patterns (confusions) are compared to human error patterns.
► The new representation made mistakes more like humans with a test-train mismatch.