Performance Implications of Frequency Domain Crossfading
for Time-Varying FIR Filtering
A common strategy to avoid audible artifacts when exchanging impulse responses used for convolution with audio signals at runtime is to use a simple crossfade. This often requires more than double the amount of operations by filtering the input signal with both impulse responses separately (often in the frequency domain), and blending the resulting signals in the time-domain. If the crossfade was applied in the frequency domain, one IFFT operation could be eliminated, at the cost of two direct convolutions between the two signal spectra with the blending functions. Most of the time however, transforming each filtered signal separately back into the time domain before blending is still significantly faster than applying the blending function through convolution in the frequency domain. With a carefully chosen blending function, direct convolution of the spectra could be faster than the time-domain multiplication. We look at why and how this technique might work and the associated performance implications.
iustin sandu
Scientific Computing Specialist
d&b audiotechnik GbmH & Co. KG
I am a software engineer at d&b audiotechnik where I work on algorithm design, implementation and optimization. I get excited about clean architecture designs and I like to think that performance does not (always) have to sacrifice elegant architecture. I have an MSc in Audio Engineering and can play a few John Mayer songs on the acoustic guitar.
