Natural signal statistics and the timecourse of spectral context effects in consonant categorization

Abstract

Speech perception is heavily influenced by surrounding sounds. When spectral properties differ between earlier (context) and later (target) sounds, this can produce spectral contrast effects (SCEs) that bias categorization of later sounds. For example, when context sounds have a low-F3 bias, listeners report more high-F3 responses to the target consonant (/d/); conversely, a high-F3 bias in context sounds produces more low-F3 responses (/g/). SCEs have been demonstrated using a variety of approaches, but most often, the context was a single sentence filtered two ways (e.g., low-F3 bias, high-F3 bias) to introduce long-term spectral properties that biased speech categorization. Here, consonant categorization (/d/-/g/) was examined following context sentences that naturally possessed desired long-term spectral properties without any filtering. Filtered sentences with equivalent spectral peaks were included as controls. For filtered context sentences, as average spectral peak magnitudes (i.e., filter gain) increased, SCE magnitudes (i.e., category boundary shifts) increased linearly. However, unfiltered context sentences showed no relationship between long-term average spectral energy and subsequent SCEs. Instead, spectral energy in the last 500 ms of unfiltered sentences predicted SCE magnitudes. Results highlight important considerations about how long-term versus short-term spectral energy in preceding sounds affects categorization of later sounds.