An online EM algorithm for source extraction using distributed microphone arrays

Expectation maximization (EM)-based clustering is applied in many recent multichannel source extraction techniques. The estimated model parameters are used to compute time-frequency masks, or estimate second order statistics (SOS) of the source signals. However, in applications with moving sources where the model parameters are time-varying, the batch EM algorithm is inapplicable. We propose an online EM-based clustering of position estimates, where the model parameters are estimated adaptively. A direct-to-diffuse ratio-based speech presence probability is used to detect noisy observations and reduce diffuse and spatially incoherent noise. The desired source signal is extracted by a multichannel Wiener filter computed using SOS estimated from the time-varying model parameters. We show that the signal of a moving source can be extracted, while reducing moving interferers and background noise.