Binaural Localization and Detection of Speakers in Complex Acoustic Scenes

Robust localisation of multiple speakers exploiting head movements and multi-conditional training of binaural cues

“…Sound source localization was performed by a Gaussian mixture model (GMM) classifier that was trained to capture the azimuth-and frequency-dependent distribution of the binaural feature space [7,8]. Given a set of K sound source directions {ϕ1, .…”

“…This can be achieved by either simulating reverberant BRIRs [7,8] or by combining HRIRs with diffuse noise [9]. In this study, binaural mixtures were created for the training stage by mixing a target source at a specified azimuth with diffuse noise, which consisted of 72 uncorrelated, white Gaussian noise sources that were placed across the full azimuth range (360…”

“…However, we note that previous computational approaches have typically been limited to locating sound sources in the frontal hemifield. Hence, although MCT has been shown to provide robust localization performance in the presence of multiple competing sources [7,8,9], the learned distribution of binaural cues for sound sources positioned in the front and rear hemifields will be quite similar. It is therefore likely that approaches that use MCT will still suffer from front-back confusions when tested under more demanding conditions.…”

“…Even using static microphones, such approaches can achieve quite promising localization performance. In order to increase the robustness of computational approaches in adverse acoustic conditions, a multi-conditional training (MCT) can be performed, in which the uncertainty of binaural cues in response to multiple sound sources and reverberation is modelled by supervised learning strategies [7,8,9]. For example, [9] report gross error rates of less than 5 % for source localization in a variety of reverberant rooms.…”

Localization of multiple sources from a binaural head in a known noisy environment

“…Binaural solutions are more difficult, and their performances drop when facing multiple sources in realistic environments (reverberation, interfering noise, etc.) [3] [4].…”