Efficient coding of multiple nonorthogonal redundancies between acoustic dimensions in novel complex sounds

Abstract

Stilp and colleagues (Proc. Natl. Acad. Sci. [2010]; JASA [2011]; PLoS One [2012]) provided perceptual evidence for efficient coding of robust covariance between acoustic dimensions in novel complex sounds. Discrimination of sounds that directly violated this redundancy (Orthogonal condition) was initially inferior to that of sounds obeying overall redundancy (Consistent condition). Performance was consistently predicted by principal components analysis (PCA), as experimental conditions aligned with statistical dimensions in PCA. Stilp and colleagues suggested efficient coding may contribute to perceptual organization for speech, but two aspects of their experimental designs qualify this extension: robust evidence for only one statistical regularity between acoustic dimensions was tested while speech possesses many; and, all statistical structure was mutually orthogonal which is often not true of speech sounds. Here, listeners discriminated sounds supporting two concurrent, nonorthogonal regularities (patterns of covariance between acoustic dimensions: attack/decay and spectral shape.) Despite nonorthogonality, these concurrent statistical regularities were efficiently coded, as discrimination of Consistent sound pairs was initially superior to that of Orthogonal sound pairs. Performance did not adhere to predictions made by PCA. Implications for speech perception and auditory ‘category’ acquisition will be discussed. [Supported by NIDCD.]