Subband Weighting for Binaural Speech Source Localization

We consider the task of speech source localization from a binaural recording using interaural time difference (ITD). A typical approach is to process binaural speech using gammatone filters and calculate frame-level ITD in each subband. The ITDs in each gammatone subband are statistically modelled using Gaussian mixture models (GMMs) for every direction during training. Given a binaural test-speech, the source is localized using maximum likelihood (ML) criterion. In this work, we propose a subband weighting scheme where subband likelihoods are weighted based on their reliability. We measure the reliability of a subband using the average frame level localization error obtained for the respective subbands. These reliability values are used as the weights for each subband likelihood prior to combining the likelihoods for ML estimation. We also introduce non-linear warping of these weights to accommodate and analyse a larger space of possible subband weights. Experiments on Subject_003 from the CIPIC database reveal that weighting the subbands is better than the unweighted scheme of combining likelihoods.