Robust probabilistic TDOA estimation in reverberant environments

In this paper, a novel Expectation-Maximization algorithm for estimating the time-delay-of-arrival of multiple non-stationary sound sources in non-stationary reverberative acoustic environments, is presented. Motivated by the success of the phase-transform/histogram based approaches of Aarabi [2, 3], the algorithm operates by learning a probabilistic relationship between the latent TDOAs and the observed microphone phase over a small collection of short-time DFTs, and in the process automatically estimates the TDOA posterior over the collection of DFTs, of each individual DFT, and also provides a measure of the frequency content of each sound source. Experimental results demonstrate that the algorithm performs as well as the Histogram techniques of Aarabi [2, 3], which have demonstrated until now unmatched results for the problem of acoustic TDOA estimation in natural environments. The model is generative an parametric and thus can potentially be seamlessly fused with probabilistic descriptions of speech production and mixing such as defined in [10], to achieve enhanced speech separation capability.