Increasing or decreasing the sampling rate of a digital audio signal is an operation included in nearly all signal processing chains, not only in digital to analogue conversion. The primary functions of upsampling are shifting the sampling images up in frequency and peak control. Changing the sampling rate requires filtering and that is where the magic happens, just like with any other filters. The following filter parameters are usually used for D-A conversion.
2x to 16x ratio
Selectable filter characteristics
Full resolution accumulator
TPDF dithering to 16/24/32 bits
Optional noise shaping
Let’s say it again, every filter is a compromise, or trade-off if you wish. It is a compromise between time and frequency domain characteristics, complexity, phase behaviour, delay, etc. There is no such thing as a perfect filter. The general conclusion is that shorter filters with less time domain distortion (ringing) and linear (or minimum) phase are sonically the least disturbing. There are many details and filter parameters that need to be separately considered for each application. Another important aspect for us is that the filter needs to be correctly designed from the technical standpoint, but also needs to sound good. That means that filters with very similar mathematical characteristics can sound dramatically different. Therefore it is crucial to evaluate all filters sonically too.
One way we do filters differently from the majority of the industry is that we don’t use zero-stuffing for oversampling filters. This method is more computationally demanding, but with the benefit of significantly reduced level of spectral images that have to be eventually removed. Samples are basically repeated (not replaced with zeros), forming a NOS-like (Non-OverSampled) waveform. Our approach makes it possible to reduce the filtering requirements and use much more relaxed filter characteristics in comparison with traditionally oversampled systems. Finally, a more relaxed filter also means less impact on audio quality.