Rossini filters questions

@AMP, can I ask you a few things about this:

The first 4 PCM filters give different trade-offs between the Nyquist image rejection and the phase response. Filter 1 has the best rejection of (unwanted) Nyquist images and the sharpest roll-off, resulting in the poorest transient response of the four. Filters 2, 3 and 4 have progressively more relaxed image rejection and progressively better transient response. Filter 2 is often preferred for orchestral music, while Filter 3 and Filter 4 are often used for rock music.
If the source data rate is 176.4, 192, 352.8 or 384kS/s, two extra filters are available. Filter 5 has a Gaussian response (which has no overshoot on transients, with a relaxed roll-off) and Filter 6 is an asymmetrical type (which features almost no pre-ringing).
There are 2 extra filters for 44.1kS/s operation also. Filter 5 is an asymmetrical design with non-linear phase and no pre-ringing. Filter 6 is a new sharp filter which has linear phase and pre-ringing. Try them and decide for yourself which you prefer.

What does this:

Filter 5 has a Gaussian response (which has no overshoot on transients, with a relaxed roll-off)


What does this:

Filter 6 is a new sharp filter which has linear phase and pre-ringing

mean? (The “sharp”.)

Why do 48, 88, 96 not have fancy filters too?


A Gaussian filter has the following properties:

  • It’s symmetrical – the impulse response will show an equal amount of energy before and after the transient)
  • It doesn’t overshoot – the impulse response won’t show the “ripples” typically seen with traditional FIR filters.
  • It has a relaxed rolloff – energy reduction in terms of dB/octave is low compared to traditional FIR filters.
  • It is short – It’s a relatively simple function that adds a minimal amount of group delay (latency)

Borrowing from Wikipedia the impulse response looks like this:

In English, it’s a relatively gentle filter that tends to smooth the signal the closer you are to the stop band. The slow rolloff means that it’s only appropriate at higher sample rates since it wouldn’t be able to suppress aliases / images sufficiently if applied close to the audioband. In other words, if applied to a 44.1k signal it wouldn’t remove enough energy before crossing the Nyquist frequency (22.05kHz). This would cause all kinds of signal artifacts in the audible and ultrasonic range and possibly let the magic smoke out of some amplifiers.

This is a more traditional FIR filter

  • The impulse response is symmetrical
  • The impulse response exhibits pre-ringing
  • It’s linear phase so its delay is applied evenly across all frequencies

The special characteristic is that it is an extremely steep filter which means that it attempts to rolloff all energy before 22.05 kHz. It’s a noble endeavor, but comes with the disadvantage of having to be a very long filter in comparison to the others. Ultimately this filter shows more pre-ringing in the time domain at a lower amplitude.

This one is for those who want a very sharp roll off (for whatever reason) and can be appealing to those who equate the number of filter taps with the quality of the filter.

Because there’s no added benefit to doing so. The Nyquist frequency is far enough away from the audio band that the more traditional filters can be implemented without the associated tradeoffs seen when the Nyquist frequency is closer to the audio band. At the same time they’re still close enough to the audio band that the Gaussian and asymmetrical filters available at the higher rates cannot be implemented without some negative side-effects.

Long story short is that there is no “one size fits all” when it comes to digital filtering. In order to accomplish the intended goal of the filter there must be compromises made in terms of the filter’s impact on the signal in the audio band. Since different people have different sensitivities to these factors we provide a choice so that each person can find the balace that’s most agreeable.