Talk:Classic Filters
Martyn wrote by email to -devel on 19Aug14
I know that 'Classic Filters' are off for 2.0.6, and don't want to revisit that.
Just a note really on Gale's comments / insistence that the algorithms in the 'Classic Filters' are 'unorthodox versus most standard documentation'. This is not true, as others have pointed out, and is a mis-reading of 'standard documentation'. Indeed the low-pass Butterworth filter in 'Classic Filters' is almost identical to the Nyquist plug-in 'Low Pass Filter', as well as being identical to the curves in Oppenheim and Schafer 'Digital Signal Processing' (my copy dating to 1975, so fairly 'classic'), all IIR filters designed with the bi-linear transform method.
The 'standard documentation' that Gale appears to be referring to is presumably http://en.wikipedia.org/wiki/Butterworth_filter which is about analogue filters, but perhaps he should be referring to something like http://en.wikipedia.org/wiki/Bilinear_transform where it talks about the effects of frequency warping for digital filters.
Further: As far as I know there is no signal-processing reason why 'Classic Filters' should not be included in a release of Audacity, but if there is a documentation issue (which there appears to be) then it shouldn't be in at this point. I am in agreement with Roger, that users should select a filter by the use of their ears.
Aside: On the manual, I don't think that we should be trying to explain the intricacies of filter design, any more than we do on http://manual.audacityteam.org/man/Equalization (which is outdated, but that is OT).
- Gale 27Aug14: No-one is suggesting we should discuss detailed intricacies of filter design. I suggest you paste the responses to Martyn here then you will see what Steve and I think are the questions the article in the Manual should address (assuming the effect does not change).
There must IMO be at least a sentence saying why the Classic Filters first order Butterworth low pass does not have a similar frequency response as the Nyquist first order low pass (despite both are digital IIR filters). We know what that sentence could say but we don't know "why" users should use Classic Filters rather than Nyquist High or Low Pass.
- Peter 28Aug14: "I suggest you paste the responses to Martyn here" - sorry can't do that, I purged my email folders yesterday so no longer have that thread. By all means delete Martyn's comment if you think it is wrong, irrelevant or not of value.
- Gale 28Aug14: Added Steve's reply and mine. Note there is a Nabble archive of -devel at http://audacity.238276.n2.nabble.com/audacity-devel-f238278.html.
- Steve on -devel: First order Classic Filters Butterworth looks (in Plot Spectrum) very distinctly different from the Nyquist version (and similarly different from the vast majority of illustrations that can be found on the Internet). http://en.wikipedia.org/wiki/Butterworth_filter looks much like the frequency response observed in the Nyquist digital filters. http://en.wikipedia.org/wiki/Bilinear_transform may talk about the effects of frequency warping for digital filters but for > 99% of our users is totally meaningless.
- Gale on -devel: The problems I see are the name of the effect, its choice of filters, and documenting
- when a user should use Classic Filters instead of Nyquist High/Low Pass?
- why is Classic Filters better than Nyquist, which might justify its 'apparently untypical' first order Low Pass? (I think we can already say why it's 'untypical')
- When should user use Chebyshev II and why?
archived ednote
- Peter 4Jan14: The name of this effect appears to be currently under discussion too, and may yet change.
- Peter 7Jan14: Martin committed the new name 'Classic Filters' in r13022 (not yet available in a nightly alpha).
- Gale 07Jan14: There is a build now, but little point replacing the image given the initialisation bug which is still unfixed. Are the controls in Butterworth type supposed to be centred or not?
- Peter 08Jan14: But are we confident that the new effect name will stick? Can we go ahead and change the name of the effect?
- Gale 08Jan14: I believe we've settled on "Classic Filter" (singular) according to the commit, but the -devel thread says "Classic Filters". I've asked on -devel. When the name is agreed there, by all means rename if you have time. I'm still catching up with the technical discussion. Thanks.
- Peter 15Jan14: Renamed this page to "Classic Filters" and fixed pages that link to here.
Some information about the filter types from Federico Miyara
"Butterworth and Chebyshev filters are polynomial filters, i.e., filters whose continuous-time attenuation is a polynomial in frequency.
Chebyshev filters are the polynomial filters that attain the highest posssible transition slope for a given order and allowable attenuation in the pass band. This means that they provide a specified selectivity at minimum cost. This property has made them very popular in analog filtering, such as antialiasing filters. Their attenuation exhibits some ripple at the pass band.
Butterworth filters are the polynomial filters that, having monotonous attenuation (no ripple) with frequency, provide the flattest frequency response in the pass band.
Chebyshev Type II filters are an intermediate between Butterworth and Chebyshev (also known as Chebysheve type I), since they have no ripple in the pass band, the same as Butterworth, buy they have higher transition slope. However, they present some "ripple" at the stop band, since the attenuation falls several times to a specified value (in the case of odd order, that value is finally reached at very high frequencies)
There are two areas were these kinds of filters they may prove useful in digital signal processing. The first one is to simulate the behaviour of the corresponding analog filter, particularly wnen investigating its transient response.
The second one is when it is necessary to implement real-time or low-latency IIR (infinite impulse response) filters. While FIR (finite impulse response) filters are capable of achieving a more accurate frequency response with low phase distortion, they usually require high orders to attain the desired selectivity, and high order implies long delay, i.e., high latency. Particularly, FIR filters based on FFT (such as the one implemented in the Equalization effect) are excellent and extremely flexible, but their computational cost is very high and they require N samples just to start yielding any output. IIR filters provide output immediately. Octave band and On third octave band filters for measurement purposes are usually implemented with IIR filters.
If post-processing audio and computation time is not an issue, do not use classic filters since the Equalization effect will provide a better result."
- Roger Dannenberg pointed out that:
- "For music/audio production, "better" is a strong word. I bet you'd be hard-pressed to hear the difference between these filters or to describe the difference. For a novice, I'd say "if one sounds better, use it; if not, don't worry, they're very similar." Engineers like to optimize things, and by choosing different things, you get different filters. But they're all doing more-or-less the same thing."
