Long-Term Temporal Audio Source Localization using SH-CRNN

Abstract

Acoustic source localization in a noisy and reverberating environment is still a challenging problem in signal processing. An improved technique has been developed herein exploring the convolutional recurrent neural network (CRNN) in the spherical harmonics domain for the far-field direction of arrival (DOA) estimation. The source signal is recorded using a spherical microphone array (SMA), and the spherical harmonics decomposition (SHD) of the recordings yields the spherical harmonics (SH) pressure coefficients. Subsequently, the SH phase and magnitude coefficients, are calculated. The CRNN model is designed and trained with long-term temporal SH magnitude and phase coefficients across all the frequencies to classify these features corresponding to the source locations. The proposed technique is assessed by extensive simulations and experimental analysis at various the signal-to-noise (SNR) ratio and reverberation time RT <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">60</inf> . The root mean square error (RMSE) is evaluated for the proposed DOA estimation technique, and a comparison with the state-of-art methods shows a significant improvement in the localization of the audio source.