Multimodal fusion for indoor sound source localization

•We propose a novel solution based on fusing visual and acoustic models to accurately identify the localization information of sound localization.•We develop a HMM-based method for separation of the acoustic transfer function (ATF) to describe clean speech sound.•We propose a new Fourier domain method for fast implementation of the HOG-type polar feature descriptor.•The proposed method has rotation-invariant capabilities and also preserves the discriminative power of extracted features.