Application of deep learning for array microphone processing

Array microphone processing is a complex application with multiple interlinked components like direction of arrival for the audio sources, beamforming and post- filtering that are dependent on the array geometry. The array microphones gained popularity by the advent of the smart speakers. In this thesis, an end-to-end solution is provided containing all of the array microphone processing components along with the denoising integrated to the core of the system using a deep learning method called autoencoders. The neural network system is trained on the magnitude spectra generated by a dataset created exclusively for this thesis by combining some of the publicly available speech and noise datasets. This thesis proposes a single channel and a multichannel speech enhancement model to solve the beamforming problem. The multichannel autoencoder model is shown to perform better than some of the common conventional beamforming methods by objective evaluation methods. Results from this thesis indicate the room for improvement in this field by the use of neural networks.