I'm working on a building a parametric EQ in Pd, and was following the model in this post, which at first appears to work really well using para_bp2~ from iemlib.

However, what I've noticed is that whenever the center frequency of a filter is set relatively low and the q approaches an octave bandwidth or so, I'm getting a lot of extraneous high frequency noise even with a gain of 0 whenever a low frequency tone is sent in to the filter. So, everything sounds well until you send a sine sweep through it and you'll notice a lot of noise when the tone is below 40 Hz or so, but again only when the center frequency of the filter is set relatively low as well.

Here is an example with very common values found on the bottom filter of a 10-band graphic EQ (31Hz, Q of 1.41), and you can hear the issue. This same issue seems to persist when using coefficients for biquad~ as well although I haven't tried that exensively.

para_bp2~issue.pd

I thought it might be aliasing caused by some ultrasonics the filter is producing for some reason, but oversampling and filtering doesn't seem to change matters much.

Anyone have any thoughts on this? Thanks for your help!

@zberkowitz said:

However, what I've noticed is that whenever the center frequency of a filter is set relatively low and the q approaches an octave bandwidth or so, I'm getting a lot of extraneous high frequency noise even with a gain of 0 whenever a low frequency tone is sent in to the filter.

It's actually quite easy to blow up a filter with a frequency near the boundary (lower or upper) and a relatively wide bandwidth.

I'm not at all surprised that you're finding a problem here. It may be a math issue rather than an implementation issue.

hjh

I thought so too, but other implementations of what I believe are the same kind of biquad formula don't seem to have this issue. 31 Hz with an octave bandwidth doesn't seem extreme!

I'm doing some testing in Supercollider on the same thing (also a biquad implementation of a peaking EQ style filter) and so far not experiencing the same issue.

Of course I'm not sure exactly how para_bp2~ is done, but using just the biquad~ object and coefficients from the ggee library I ended up with the same result.

@zberkowitz said:

I thought so too, but other implementations of what I believe are the same kind of biquad formula don't seem to have this issue. 31 Hz with an octave bandwidth doesn't seem extreme!

I'm doing some testing in Supercollider on the same thing (also a biquad implementation of a peaking EQ style filter) and so far not experiencing the same issue.

One possible difference is that SC filters store coefficients as double -- meaning (I think) that the coefficient multiplication/addition is also double. Signal ins/outs are single-precision in both SC and Pd, but if the internal math is double in SC and single in Pd (a quick glance at the Pd source suggests that this is the case), that would likely translate into lower noise in the SC filters.

Maybe?

hjh

@ddw_music I don't know enough to know if that could cause what I'm hearing but interesting idea! Not much to do about that unfortunately if it is the case...

@ddw_music You could be on to something! https://ez.analog.com/dsp/sigmadsp/w/documents/5196/when-should-i-use-single-precision-or-double-precision-filters