Neural network-based adaptive noise cancellation for enhancement of speech auditory brainstem responses

The measurement of the speech-evoked auditory brainstem response (speech ABR) is a promising technique for evaluating auditory function. However, the speech ABR is severely contaminated by background noise related to other brain electrical activity. The most commonly used method to enhance the signal-to-noise ratio (SNR) of the response is coherent averaging, while recently adaptive filtering has also been reported. All of the applied methods are based on linear operations, but since the assumption of linearity may not be valid for neural activity, linear methods may not be adequate. In this paper, we present a new nonlinear adaptive noise cancellation (ANC) based on a multilayer perceptron neural network to enhance the speech ABR and compare its performance with a linear ANC algorithm based on least mean squares adaptive filtering. The effectiveness of the methods is tested using speech ABR data and is based on two different types of SNR measures, the local SNR at the fundamental frequency of the response and the overall SNR. The results show that the nonlinear neural network-based ANC can reduce the required recording time and performs better than the linear ANC especially when the SNR of the recorded speech ABR is low.