Forums

IIR audio filterbank

Started by Robert Adams October 12, 2010
I am looking for a way to design an audio IIR filterbank. I need to
split an audio signal into roughly 1/3 octave bands with 6th-order or
8th-order IIR bandpass filters and then sum them back together again
with a ripple of +/- 1 dB. I can do a "seat-of-the-pants" design that
works well until I approach Nyquist where it kind of falls apart. I
have very limited MIPs so I can't do anything very fancy. Any
pointers?


Bob

Robert Adams wrote:

> I am looking for a way to design an audio IIR filterbank. I need to > split an audio signal into roughly 1/3 octave bands with 6th-order or > 8th-order IIR bandpass filters and then sum them back together again > with a ripple of +/- 1 dB. I can do a "seat-of-the-pants" design that > works well until I approach Nyquist where it kind of falls apart. I > have very limited MIPs so I can't do anything very fancy. Any > pointers?
Constant Q transform? Vladimir Vassilevsky DSP and Mixed Signal Design Consultant http://www.abvolt.com
On Oct 12, 9:37=A0pm, Vladimir Vassilevsky <nos...@nowhere.com> wrote:
> Robert Adams wrote: > > I am looking for a way to design an audio IIR filterbank. I need to > > split an audio signal into roughly 1/3 octave bands with 6th-order or > > 8th-order IIR bandpass filters and then sum them back together again > > with a ripple of +/- 1 dB. I can do a "seat-of-the-pants" design that > > works well until I approach Nyquist where it kind of falls apart. I > > have very limited MIPs so I can't do anything very fancy. Any > > pointers? > > Constant Q transform? > > Vladimir Vassilevsky > DSP and Mixed Signal Design Consultanthttp://www.abvolt.com
Sounds right, but it needs to sum directly to create an approximation of the input; that is, I can't afford a "synthesis filterbank".

Robert Adams wrote:

> On Oct 12, 9:37 pm, Vladimir Vassilevsky <nos...@nowhere.com> wrote: > >>Robert Adams wrote: >> >>>I am looking for a way to design an audio IIR filterbank. I need to >>>split an audio signal into roughly 1/3 octave bands with 6th-order or >>>8th-order IIR bandpass filters and then sum them back together again >>>with a ripple of +/- 1 dB. I can do a "seat-of-the-pants" design that >>>works well until I approach Nyquist where it kind of falls apart. I >>>have very limited MIPs so I can't do anything very fancy. Any >>>pointers? >> >>Constant Q transform? >> > Sounds right, but it needs to sum directly to create an approximation > of the input; that is, I can't afford a "synthesis filterbank".
What is the application ? I.e. if this is supposed to be a graphic equalizer, there are many more efficient ways to accomplish that. Vladimir Vassilevsky DSP and Mixed Signal Design Consultant http://www.abvolt.com

Robert Adams wrote:

> On Oct 12, 9:37 pm, Vladimir Vassilevsky <nos...@nowhere.com> wrote: > >>Robert Adams wrote: >> >>>I am looking for a way to design an audio IIR filterbank. I need to >>>split an audio signal into roughly 1/3 octave bands with 6th-order or >>>8th-order IIR bandpass filters and then sum them back together again >>>with a ripple of +/- 1 dB. I can do a "seat-of-the-pants" design that >>>works well until I approach Nyquist where it kind of falls apart. I >>>have very limited MIPs so I can't do anything very fancy. Any >>>pointers? >> >>Constant Q transform?
> Sounds right, but it needs to sum directly to create an approximation > of the input; that is, I can't afford a "synthesis filterbank".
P.S. You do know of Linkwitz-Riley filters, do you? VLV
On Oct 12, 11:59=A0pm, Vladimir Vassilevsky <nos...@nowhere.com> wrote:
> > P.S. You do know of Linkwitz-Riley filters, do you? >
you might find both Bob's name and Siegfried Linkwitz listed here: http://www.aes.org/awards/ . (i think Bob knows about L-R). r b-j
On Oct 13, 12:10=A0am, robert bristow-johnson
<r...@audioimagination.com> wrote:
> On Oct 12, 11:59=A0pm, Vladimir Vassilevsky <nos...@nowhere.com> wrote: > > > > > P.S. You do know of Linkwitz-Riley filters, do you? > > you might find both Bob's name and Siegfried Linkwitz listed here:http://=
www.aes.org/awards/.
>
fuck! i can't believe it. i *knew* i would see Bob's name on the list because i know a couple of times he got AES awards. but i cannot believe that Linkwitz's name is not on the list. unbelievable! the AES never awarded Linkwitz and Riley a fellowship or, at least, a publication award???!! can't figger that one out. r b-j
On Oct 13, 12:12=A0am, robert bristow-johnson
<r...@audioimagination.com> wrote:
> On Oct 13, 12:10=A0am, robert bristow-johnson > > <r...@audioimagination.com> wrote: > > On Oct 12, 11:59=A0pm, Vladimir Vassilevsky <nos...@nowhere.com> wrote: > > > > P.S. You do know of Linkwitz-Riley filters, do you? > > > you might find both Bob's name and Siegfried Linkwitz listed here:http:=
//www.aes.org/awards/.
> > fuck! =A0i can't believe it. =A0i *knew* i would see Bob's name on the > list because i know a couple of times he got AES awards. =A0but i cannot > believe that Linkwitz's name is not on the list. =A0unbelievable! =A0the > AES never awarded Linkwitz and Riley a fellowship or, at least, a > publication award???!! > > can't figger that one out. > > r b-j
Yes I know about L-R, but I am looking for something with > 30 bandpass filters, not sure L-R can be extended to that case? Bob
On Oct 13, 4:37=A0am, Robert Adams <robert.ad...@analog.com> wrote:
> On Oct 13, 12:12=A0am, robert bristow-johnson > > > > > > <r...@audioimagination.com> wrote: > > On Oct 13, 12:10=A0am, robert bristow-johnson > > > <r...@audioimagination.com> wrote: > > > On Oct 12, 11:59=A0pm, Vladimir Vassilevsky <nos...@nowhere.com> wrot=
e:
> > > > > P.S. You do know of Linkwitz-Riley filters, do you? > > > > you might find both Bob's name and Siegfried Linkwitz listed here:htt=
p://www.aes.org/awards/.
> > > fuck! =A0i can't believe it. =A0i *knew* i would see Bob's name on the > > list because i know a couple of times he got AES awards. =A0but i canno=
t
> > believe that Linkwitz's name is not on the list. =A0unbelievable! =A0th=
e
> > AES never awarded Linkwitz and Riley a fellowship or, at least, a > > publication award???!! > > > can't figger that one out. > > > r b-j > > Yes I know about L-R, but I am looking for something with > 30 > bandpass filters, not sure L-R can be extended to that case? > > Bob- Hide quoted text - > > - Show quoted text -
Application is a bit like a graphic EQ but I need to measure the energy in each band with pretty good selectivity, and my MIPs are so limited that I wanted to use the same filters for energy estimation as for the graphic EQ, which means I need to use the "parallel-sum" form of graphic EQ instead of the "cascade-of-peaking-filters" form. Unfortunately this means you get all the messy phase effects when each filter crosses over with its upper and lower neighbors. Bob
On Oct 13, 4:42=A0am, Robert Adams <robert.ad...@analog.com> wrote:
...
> Application is a bit like a graphic EQ but I need to measure the > energy in each band with pretty good selectivity, and my MIPs are so > limited that I wanted to use the same filters for energy estimation as > for the graphic EQ, which means I need to use the "parallel-sum" form > of graphic EQ instead of the "cascade-of-peaking-filters" form. > Unfortunately this means you get all the messy phase effects when each > filter crosses over with its upper and lower neighbors.
Bob, i might have mentioned this to you one of those times we were having lunch at the Green Papaya, but an idea for a paper that i had but have done little about is an automated way of turning that "cascade-of-peaking-filters" form into an equivalent parallel-sum form: a graphic EQ with 3 sliders per octave, and a pretty constant 1/3 octave bandwidth on each peaking filter. but then apply (automagically) the Heaviside partial fraction expansion to this thing and come out with a bunch of parallel 2nd-order filters (each with 2 coefs in the numerator) and one feedforward path of the dry input. i sorta think this would be a good thing for an EE grad student to do. the reason i liked it this way is because the double-precision word that results from every parallel filter need not be truncated, but summed in a big accumulator with that result truncated (and noise- shaped). being parallel, it might be cleaner than 25 peaking-EQs in series, each quantizing its output. but being the same transfer function as the 25 peaking-EQs in a row, it wouldn't have the phase cancellation problem of the side skirts of adjacent bands typical in the parallel context. r b-j