Active 2D Sound Source Localization using a Sample DCNN Architecture & Encoding
Niraj Anil Babar, J Sumanth, Noel Alben, Neeluru Sai Deepika, S.N.R Ajey, Mr J. P. Pramod · 2022 2nd International Conference on Intelligent Technologies (CONIT) · 2022
This paper presents how to use raw audio data (multi-channel) to train a deep convolutional neural network in an end-to-end architecture to locate active auditory sources in space. It employs a revolutionary Sample DCNN architecture and encoding approach to encode the spatial coordinates of sources, allowing for end-to-end 2D source localisation. A Transfer Learning approach is employed to train the DCNN. Using clean voice data from the TIMIT database's dialect 8, we synthesise and produce anechoic and reverberant datasets. We map the model's predictions on a 2D plane to visualise them.