An event-driven probabilistic model of sound source localization using cochlea spikes

Abstract

This work presents a probabilistic model that estimates the location of sound sources using the output spikes of a silicon cochlea such as the Dynamic Audio Sensor. Unlike previous work which estimated the source locations directly from the interaural time differences (ITDs) extracted from the timing of the cochlea spikes, the spikes are used instead to support a distribution model of the ITDs representing possible locations of sound sources. Results on noisy single speaker recordings show average accuracies of approximately 80% on detecting the correct source locations and an estimation lag of <;100ms.

Speaker Activity Detection and Minimum Variance Beamforming for Source Separation

“…The second is M-NICA introduced for blind source separation (BSS) of source envelopes. The third is a recently proposed localization method that uses the timing of asynchronous samples from an event-driven binaural audio sensor [12]. We considered this third method because of its possible low computational complexity.…”

“…It was recently shown how the outputs of this event-based cochlea sensor could be used to localize multiple active sources simultaneously [12]. Each output event of the sensor is assigned a probability of it being produced by a source at a particular location l ∈ L = {1, ...L} where L is the number of possible locations.…”

“…The event count feature corresponds to a moving average of events collected in a defined time window. We use this feature as an estimate of the speech envelope [12]. Once an envelope estimate is obtained for each location, the locations of the active sources have to be selected.…”

“…We extend the work in [12] by using the estimated envelopes in a SAD system. By comparing the various envelopes using (1), we construct a SAD similar to the M-NICA method.…”

“…This work proposes such a framework with the only assumption of knowing the number of speakers. First, speaker activity over short time frames of 20ms is estimated using different algorithms such as GCC-PHAT [10], multiplicative nonnegative independent components analysis (M-NICA) [11] and spike separation (SPS) [12]. From the results of the SAD, time frames assigned to one speaker are pooled together and the ATFs are then estimated using a noise-covariance whitening based ATF estimation method [6].…”