Low-latency real-time blind source separation for hearing aids based on time-domain implementation of online independent vector analysis with truncation of non-causal components

In this paper, we present a low-latency scheme for real-time blind source separation (BSS) based on online auxiliary-function-based independent vector analysis (AuxIVA). In many real-time audio applications, especially hearing aids, low latency is highly desirable. Conventional frequency-domain BSS methods suffer from a delay caused by frame analysis. To reduce the delay, we implement separation filters as multiple FIR filters in the time domain, which are converted from demixing matrices estimated by online AuxIVA in the frequency domain. Also, to further reduce the latency, part of the non-causal components of the FIR filters are truncated on the basis of causality analysis for ideal separation filters using a simple model. By experimental evaluation using a head and torso simulator in a real environment, the proposed algorithm with an algorithmic delay of less than 10 ms exhibited a separation performance of 7.7 dB in terms of the signal-to-interference ratio (SIR), which was less than 1.4 dB degradation from the case of conventional frequency-domain implementation.