Binaural Sound Source Localisation is increasingly being achieved by means of the Convolutional Neural Network (CNN). These networks take in a Time-Frequency representation of audio as an input, and use this to estimate the direction of arrival of a sound. In previous works, different Time-Frequency representations have been used, but never only using solely magnitude spectra, leading to a lack of understanding in the importance of this in full azimuthal binaural sound source localisation. This work aims to address that gap by testing the performance of a CNN trained and tested on four different Time-Frequency representations: Mel-Spectrogram, Gammatonegram, Mel-Frequency Cepstrum, and Gammatone-Frequency Cepstrum. From this test, it was found that Spectrograms are suitable for the task of full azimuthal binaural sound source localisation.
(See document for exact affiliation information.)
https://aes2.org/publications/elibrary-page/?id=22084
(302KB)
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member Join the AES. If you need to check your member status, login to the Member Portal.
