Efficient frequency-domain filter crossfading for fast convolution with application to binaural synthesis
Binaural synthesis of complex, dynamic acoustic scenes requires multiple convolution processes with time-varying head-related transfer functions (HRTFs). To avoid audible artifacts, this requires not only a position-dependent interpolation of HRTFs, but also smooth transitions during filter changes. Existing solutions either use gradual interpolation of filter coefficients or crossfading of filter outputs. This paper proposes a crossfading method that operates in the frequency domain, which enables integration with fast frequency-domain techniques such as overlap&save, uniformly or non-uniformly partitioned convolution. Fading is performed by circular convolution with a frequency-domain window function. A method is proposed to design such windows with only a few non-zero coefficients, resulting in significant complexity reductions compared to existing crossfading techniques. Further substantial efficiency improvements are achieved by integrating the proposed crossfading method into binaural rendering systems.