Section: Application Domains

Sound-source separation and localization

We explore the potential of binaural audition in conjunction with modern machine learning methods in order to address the problems of sound source separation and localization. We exploit the spectral properties of interaural cues, namely the interaural level difference (ILD) and the interaural phase difference (IPD). We have started to develop a novel supervised framework based on a training stage. During this stage, a sound source emits a broadband random signal which is perceived by a microphone pair embedded into a dummy head with a human-like head related transfer function (HRTF). The source emits from a location parameterized by azimuth and elevation. Hence, a mapping between a high-dimensional interaural spectral representation and a low-dimensional manifold can be estimated from these training data. This allows the development of various single-source localization methods as well as multiple-source separation and localization methods.