Languages across the world are efficiently coded by the auditory system

Abstract

Independent Component Analysis (ICA) is a powerful method for uncovering statistical structure in natural stimuli. Lewicki (2002 Nature Neuroscience) used ICA to examine statistical properties of human speech. Filters that optimally encoded speech were an excellent match for frequency tuning in the cat auditory nerve, leading to suggestions that speech makes efficient use of coding properties in the mammalian auditory system. However, Lewicki only examined American English, which is neither normative nor representative of the world's languages. Here, fifteen languages were examined (Dutch, Flemish, Greek, Javanese, Ju|'hoan, Mandarin Chinese, Norwegian, Swedish, Tagalog, Tahitian, Urhobo, Vietnamese, Wari', Xhosa, Yeyi). Each recording contained speech tokens from native speakers without any background noise for at least seven minutes. Maximum likelihood ICA was used to create statistically optimal filters for encoding sounds from each language. These filters were then compared to the same physiological measures analyzed in Lewicki (2002). Languages produced a range of ICA solutions, as expected, but were highly consistent with both statistically optimal filters for American English and physiological measures. Results significantly extend Lewicki (2002) by revealing agreement between response properties of the auditory system and speech sounds in a wide range of languages.