A friend asked which filter setting he should use for the DAC [digital‑to‑analogue converter] of his music streamer. I wasn’t aware that selectable filter characteristics for DACs even existed; I’d assumed that the whole function of a DAC filter was to do with the Nyquist theorem and aliasing, and eliminating undesirable artefacts entering the audible spectrum. If so, why would you want to play around with filtering characteristics? Surely all of this takes part outside the audible frequency spectrum (20Hz to 20kHz). Or have I misunderstood?
SOS Forum post
SOS Technical Editor Hugh Robjohns replies: Several D‑A and some A‑D converter‑chip manufacturers offer selectable filter characteristics, and some product manufacturers extend that selection capability to the user. The filters sound slightly different with some material, predominately that with a lot of energy at very high frequencies, because most options impinge on the very top of the audio band.
The vast majority of converter filters are so‑called ‘half‑band’ filters, and although their response is flat up to 21 or 22kHz, they only manage 6dB of attenuation at the Nyquist frequency.
In theory, the D‑A reconstruction filter should have a brickwall response just below the Nyquist frequency (half the sample rate). So in a 44.1kHz system it should be flat up to 21 or 22kHz and offer upwards of 100dB attenuation at 22.05kHz (the Nyquist frequency) and above. But this is close to impossible and largely impractical to achieve in hardware, and the vast majority of converters use so‑called ‘half‑band’ filters which are much easier to build. Although their response is flat up to 21 or 22 kHz, half‑band filters manage only 6dB of attenuation at the Nyquist frequency, but there’s usually so little audio around the Nyquist frequency that the aliasing resulting from only 6dB of attenuation is negligible with most music. This implementation is often called a ‘fast’ filter. An alternative, often called a ‘slow filter’, is to start the filter roll‑off slightly earlier, to a much better attenuation at the Nyquist frequency. Obviously, that curtails the audible HF response around 20kHz very slightly, but it removes aliasing artefacts with some material, and some people prefer that particular compromise.
Another consideration is the time‑domain response. The linear‑phase filters traditionally employed preserve the phase relationships between harmonics across the audio bandwidth and allow incredibly steep filter cutoffs, which is desirable. But they also introduce pre‑ringing artefacts which don’t exist in the natural world. As some people believe they can hear this pre‑ringing, many converter chips include one or more ‘minimum‑phase’ filter options. These replicate analogue filters and have only (natural) post‑ringing artefacts. Sadly, they also have much gentler filter slopes, so the compromise here involves even greater HF roll‑off, or more aliasing, or both.
In general, although a 44.1kHz digital system has a nominal 20kHz bandwidth, few people can hear sounds that high so most are willing to sacrifice a little of their inaudible HF extension to avoid both audible aliasing and pre‑ringing.