Fast Random Approximation of Multi-Channel Room Impulse Response

The training of modern neural-network-based speech processing systems typically requires a large amount of reverberant data to make the systems robust against reverberation. Existing room impulse response (RIR) filter simulation tools have proven effective in a wide range of speech processing tasks and neural network architectures, but their usage in on-the-fly data simulation pipeline still remains questionable. In this paper, we propose FRAM-RIR, a fast random approximation method of the widely-used image-source method (ISM), to efficiently generate multi-channel RIR filters used for data augmentation. FRAM-RIR is an extension of our previously proposed FRA-RIR method for multi-channel scenario, which maintains accurate direction-of-arrival (DOA) information of all sound sources in a space. Visualization of oracle beampatterns shows that FRAM-RIR can generate more realistic RIR filters than existing widely-used ISM-based tools, and experiment results on multi-channel noisy speech separation task show that models trained with FRAM-RIR can also achieve on par or better performance with a significantly accelerated training procedure. We release an implementation of FRAM-RIR online 1.