OBFNet: A Neural Beamforming Network Using Orthonormal Basis Functions

Time-domain neural beamformers, such as the filter-and-sum network (FaSNet), can alleviate residual noise and reduce algorithmic latency compared with short-time Fourier transform domain beamformers. However, as a finite impulse response (FIR) beamformer, FaSNet requires a long filter length to achieve high performance. In this work, we therefore propose a time-domain neural beamformer using orthonormal basis functions (OBFs), called OBFNet. We incorporate an OBF module to reduce the filter length while maintaining the filtering gain. Because OBF filters have adjustable poles, OBFNet is more flexible and can be regarded as a generalization of FaSNet. The OBF filter coefficients are optimized in an all-neural form, where a set of learnable parameters is trained for pole selection and a sequential learning network is used to estimate the adaptive weights. We present the impulse response (IR) calculation of OBFNet and perform the standard filter-and-sum operation to recover the target signal. The OBFNet is trained using multi-task learning, where an additional frame-level loss is introduced to constrain the IR. Differences between the IRs of OBFNet and FaSNet can thus be inspected through the loss, which also affects the speech quality. Experimental results show that the proposed OBFNet can outperform existing methods even with a shorter filter length, which is helpful to reduce the latency.