Abstract
A novel system of computer algorithms is formulated to perform onset-guided source localisation using binaural stimuli. This system, called the spatial analyser, will analyse spatial attributes including source location. It is computationally efficient, compatible with streamed binaural data, and uses psychophysically-valid analysis techniques wherever possible. The main components of the system are a model of the human auditory periphery, an onset detector, a running localisation algorithm, and some logic to combine these. The onset detector is designed specifically for spatial analysis, using a combination of linear regression and band-pass filtering techniques to produce a response that is sensitive to auditory onsets and robust to noise. It also features an implementation of the precedence effect. To localise sounds, an efficient method is found for extracting interaural time difference cues using the interaural cross-correlation function. Instantaneous interaural time and intensity differences of the binaural signal are calculated and mapped to lateral angle using a database of interaural cues. A cross-weighting formula combines the interaural time and intensity data across frequency bands. Loudness weighting is then applied to every critical band to produce an output. Spatial information is handled throughout the localisation algorithm in the form of lateral angle histograms. These are discrete functions, which specify localisation strength against lateral angle for any particular combination of cues. In a series of validation experiments, the spatial analyser determines the direction of most sound sources to within 10 in a reverberant environment. For most sources, this performance is maintained even when a substantial amount of white noise is added to the audio as a confusing signal. The output data is also shown to be compatible with auditory source width extraction. With slight modifications, the spatial analyser can also approximate source distance.