Importance of time-frequency units for sentence recognition as a function of signal-to-noise ratio

Speech recognition in noise may be viewed as a classification problem, in which the auditory system selects a subset of time-frequency (T-F) units to build a representation of the signal. The present study introduces local signal-to-noise ratio (SNR) importance functions, an approach to determine the contribution to speech intelligibility of T-F units as a function of their SNR. Consistent with previous work from this group on auditory-channel independence, it was hypothesized that T-F units with a relatively equal mixture of signal and noise (around 0 dB SNR) may be the most disruptive for intelligibility because they should be the most difficult to classify. This hypothesis was assessed by measuring the intelligibility of sentences presented in a competing-speech background while discarding a fixed proportion of units all with the same local SNR. The stimuli were either vocoded or left unprocessed to evaluate the influence of fine-structure cues on classification and also help better understand why coch...