Acoustic Source Localization in the Circular Harmonic Domain Using Deep Learning Architecture

The problem of direction arrival (DOA) estima- tion with a circular microphone array has been addressed classical source localization methods, such as the model-based methods and parametric methods. These have an ad- vantage in estimating DOAs blind manner, i.e. no (or limited) prior knowledge about sound sources. However, their performance tends to degrade rapidly noisy reverberant environments or presence sensor limitations, gain phase errors. In this paper, we present new approach by leveraging strength convolutional neural network (CNN)-based deep learning approach. particular, design harmonic features that are frequency- invariant inputs CNN architecture, so offer improvements DOA estimation unseen adverse obtain good adaptation imperfections. To our knowledge, not used domain. Experiments performed on both simulated real-data show method gives significantly better performance, than recent baseline variety noise reverberation levels, terms accuracy estimation.