CRNN vs. Self-Attention in Source Localization and an Introduction of the Project HAIKUS

The seminar will consist of two parts, the first part will be related to DOA estimation work done during my master internship, and the latter part will describe some preliminary results about the project HAIKUS. Recently, RNN based CRNN architecture made the state of the art success with audio signals in DOA estimation and SELD tasks, but they are too costly to train and difficult to be used in embedded applications. So can we take advantage of the transformers based self-attention modules to replace LSTM and RNN and help the model learn faster with fewer parameters and make the system parallelizable? Project haikus is all about machine learning for audio augmented reality, this seminar will focus upon some of my early results in acoustic parameter estimation, which is one of the key point for virtual rendering of the sound source in a user environment.