Robust decoding of the speech envelope from EEG recordings through deep neural networks

During speech perception, a listener's brain activity tracks amplitude modulations in the speech signal, which are encoded in the speech envelope. The neural tracking of the speech envelope is modulated by cognitive factors such as attention to one of several competing speakers. Decoding the speech envelope from noninvasive electroencephalography (EEG) may be useful in future auditory prosthesis that could restore speech comprehension in noisy environments. Such applications require, however, a robust decoding of the speech envelope that functions in different acoustic conditions and that generalizes between different participants. Here we show that deep neural networks (DNNs) can lead to an enhanced decoding that has around 38% higher performance than the standard method, linear regression. The advantage of the DNNs persists across different acoustic scenarios and also when listener-independent decoders are used. We also show how an improved envelope decoding performance translates into a higher auditory attention decoding accuracy for the DNNs, in comparison to the method of linear modelling. Our work therefore demonstrates that DNNs show promise for data efficient, real-world auditory attention decoding.