Noise robust representation of speech in the primary auditory cortex.

Abstract

It is well known that humans can robustly perceive phonemes despite substantial variability across speakers, context and natural distortions. This study examines the responses of neurons in primary auditory cortex (A1) to phonetically labeled speech stimuli in clean, additive noise and reverberant conditions. Using a linear decoder [Bialek (1991)] to reconstruct the input stimulus spectrogram from the population response, we observed that spectrograms reconstructed from the neural responses to noisy speech were closer to the original clean spectrograms than to the noisy ones. This indicates that sound representations in A1 serve to enhance information about natural speech signals relative to noise, thus extracting signal from noise. Examining the average reconstructed phoneme spectrograms in clean and noisy speech revealed a remarkable robustness in the encoding of features important for discrimination of different phonemes. In addition, it was found that the strict linear spectro‐temporal receptive field (STRF) model of A1 neurons is insufficient to explain the noise robustness observed in the neural data. However, when a non‐linear synaptic depression is integrated into the inputs for the STRF model, the noise was reduced in the reconstructed spectrograms similar to what observed with the actual neural data.