Project: Filter Calculator for Convolution

[jdc0589]

I put this in the PE forums as well.

Background:
Some of you may remember my 3-CR Full Range project where I created a filter in REW and used a convolution plugin with Foobar audio player to implement it.

For those of you that don't know, convolution is similar to an 'infinite' point EQ. My project was a full range one, so I took measurements with NO signal manipulation of any kind from physical components (notch filters, bsc, etc..) or from an EQ. From those raw measurements, which generate a standard frequency response graph (an ugly one), you calculate the opposing curve that would balance your F/R. The opposing curve is used in the convolution process to generate flat frequency response.

New Stuff:
The automatic calculation functions in REW....suck. They are limited in the frequency range in which they work and how many peaks/valleys can be addressed. So, last week I wrote a utility that reads raw frequency response data from a text file, calculates the target Db level, generates the opposing curve, and saves the data to another text file. It needs some more work in regard to target Db calculation/optimization and point density in relation to efficiency, but the initial results look pretty good.

My goal is to expand the utility over this comming summer to output the final .wav file used by convolver plugins so no extra software/step is required before the output is ready for convolution.
All code is in Java, so it will be able to run on pretty much any platform, regardless of operating system. If you have any interest or knowledge in this type of signal processing, please let me know, Im learning as I go here.


Thanks for looking! If I can keep developing and testing the program to the point I am completely satisfied with it, I will provide anyone interested with a copy.

Post any questions or comments!

[JonMarsh]

very interesting- not something I have background knowledge in, but something I may have need of in the next few months- what would be interesting is also having a tabular output of the correction equalization- poles and zeros and levels.

[jdc0589]

very interesting- not something I have background knowledge in, but something I may have need of in the next few months- what would be interesting is also having a tabular output of the correction equalization- poles and zeros and levels.

It would be very simple to add the ability for it to output the +/- Db value at each of the measurement points


Also, If anyone has FR data from a project in text form, I could really use some more test data. Optimally this data would have been taken WITHOUT any signal filtering (via physical passive filters or eq), IE: it would look nastier than the result you got after implementing filter networks. If it was a multi-way project, obviously its ok if the crossover (without any filters) was in place during measurements.

In addition, if you have F/R data for a finished project (with passive filter networks implemented in the crossover) and just want to see what the difference would be with convolution applied on top of your passive stuff, I'll run it for you and post a graph :T .

[Hank]

This sounds interesting (and over my head, but I would be interested in the end result). Mark "The Kid" my have knowledge you are seeking. I'll ping him. I assume this could work for someone to determine the numbers for somethink like, oh, say, an open baffle line array with digital filters? :W

[jdc0589]

Ill put an example up. The calculation is quite simple currently seeing as I have not taken point density vs. efficiency in to account yet (by that I mean, If you need a tv with 1600x1020 resolution to get the job done, having 1900x1200 would be inefficient).
I also may change it up so there is NO signal boost at any point, meaning lower target output would be used with padding everywhere (to avoid going above digital 0).

[jdc0589]

This sounds interesting (and over my head, but I would be interested in the end result). Mark "The Kid" my have knowledge you are seeking. I'll ping him. I assume this could work for someone to determine the numbers for some think like, oh, say, an open baffle line array with digital filters? :W

Here is the filter curve for some of my test data. The RAW frequency response is essentially the inverse of this graph, centered around the same target Db point (I think it was 62 Db here). Final F/R should theoretically look like the average between the filter graph and Raw response graph, which would be a flat line.

I realize the X-axis tick marks are on drugs, but the scale is 20-20000hz.

[ch83575]

Have you seen this: http://www.acourate.com/ ?

There are a few video tutorials on how to use the software that give great insight into the uses of the various functions...

Very interesting stuff this. Keep up the good work.

-Chad

[jdc0589]

Have you seen this: http://www.acourate.com/ ?

There are a few video tutorials on how to use the software that give great insight into the uses of the various functions...

Very interesting stuff this. Keep up the good work.

-Chad

I had written it off due to the $500 price tag. But, I may download the trial and see if the functions I need are enabled.

[ch83575]

Hey, don't get me wrong... I am all for DIY, even software. If you can come up with a utility to accomplish what needs to be accomplished to implement this without the $500 software that is great! Especially if it is something you might plan on sharing with the rest of us when you are done ;x( . I do think the acourate white paper is a good read and the video tutorials give you an idea of how the software works. I was just pointing it out so you might not have to re-invent the wheel in every aspect. There are also interesting conclusions I have seen related to the acourate software that will probably hold true with your own endeavors; for instance, I have read on the acourate website and in a build thread where the software was used that using the FIR for driver linearization outside of the passband can actually be harmful to the finished sound even though it makes it follow the target more closely.

All just food for thought. I for one am very interested in your project, so keep us posted.

-Chad

[ergo]

Take a look at

http://www.hypex.nl/docs/Bruno%20Masterclass/slides.htm

From slide 106 onwards this presentation has some interesting info on software equalizing the loudspeakers. (I did not figure out a way to link to that page straight... so one has to look from beginning onwards, but that is interesting as well anyhow)

My experience this far is similar, ie it is very worthwhile to measure at more than one point, average the responses and then try to compensate only the peaks or dips that show up in the average. How far off axis one should go from on axis is of course also a big question.

Ergo

[ch83575]

Take a look at

http://www.hypex.nl/docs/Bruno%20Masterclass/slides.htm

From slide 106 onwards this presentation has some interesting info on software equalizing the loudspeakers. (I did not figure out a way to link to that page straight... so one has to look from beginning onwards, but that is interesting as well anyhow)

My experience this far is similar, ie it is very worthwhile to measure at more than one point, average the responses and then try to compensate only the peaks or dips that show up in the average. How far off axis one should go from on axis is of course also a big question.

Ergo

What a great presentation. Good read all the way through. Talk about a totally different take on the use of FIR filters in loudspeakers. Dr. Brüggemann (who wrote the acourate software) seems to think the subtractive filter generation technique is the holy grail of loudspeaker design, and Bruno Putzeys (the slideshow guy... apparently works for Hypex) says that the technique is... and I quote: "naive."

It does seem to me that canceling pre-ringing with more pre-ringing in non-coincident drivers would be dubious at best.

Of course this thread is supposed to be about driver linearization using convolution, which is also discussed at some length. Again, very cool read... thanks for the link!

jdc0589,
Are you using a passive crossover in the speaker? The convolution is at the source and for the entire system response?

-Chad

[TacoD]

What a great presentation. Good read all the way through. Talk about a totally different take on the use of FIR filters in loudspeakers. Dr. Brüggemann (who wrote the acourate software) seems to think the subtractive filter generation technique is the holy grail of loudspeaker design, and Bruno Putzeys (the slideshow guy... apparently works for Hypex) says that the technique is... and I quote: "naive."

It does seem to me that canceling pre-ringing with more pre-ringing in non-coincident drivers would be dubious at best.

Of course this thread is supposed to be about driver linearization using convolution, which is also discussed at some length. Again, very cool read... thanks for the link!

jdc0589,
Are you using a passive crossover in the speaker? The convolution is at the source and for the entire system response?

-Chad

Bruno designed the ucd switching amplifiers, so I think he knows a thing or two about filters.

[jdc0589]

jdc0589,
Are you using a passive crossover in the speaker? The convolution is at the source and for the entire system response?

-Chad

Currently, there is nothing passive at all; Its a straight shot from the amp to the speakers. But, Ill go ahead and say its probably a better idea to do a little general filtering via passive components and do moderate/fine tuning via corrective software; especially if its a multi-way, I would lean towards more passive solutions around crossover points.

I e-mailed the guys at acourate and they said it would easily be able to take my text based f/r curves and turn them in to filter files. I may see if the demo will let me.

[Amphiprion]

Hi,

Like Hank said, I have some experience with this. My senior design project in college was a DSP based loudspeaker system. We measured each individual driver with Liberty Instruments LAUD, imported the resulting CSV files into Matlab, turned the CSV files into complex frequency responses for each driver. Then we divided an idealized frequency response for each driver by its actual complex frequency response. The result was a complex response that was perfect for correcting the drivers but was not describable with simple polynomial based filters. We approximated these idealized correction plots with a built-in function in Matlab (invfreqz) to turn them into IIR filters for each driver that would essentially be a perfect crossover to get each driver's raw response to match what we wanted that driver's final response to be.

This was six or seven years ago, and I can dig up the work if you want. It was fairly successful, but we only used ~10th order polynomials to approximate the 'ideal' transfer function for each driver. Matlab got unreliable in doing higher-order systems, probably something do to with precision problems with the built in algorithm.

[davey_m]

You could try the open source package Octave too. Much commonality with Matlab and it has all the functionality you need.

With Octave it is easy to derive filter, driver correction, and combined driver correction and filter functions.

What makes it a little more involved are the following two issues:

- With driver correction functions you usually want to limit correction for sharp deep dips, as it might increase the required amplfier voltage swing too much.

- When using FIR filters you must also check for pre-ringing, which might be audible. A certain amount of pre-ringing is inaudible and therefore acceptible. Use accepted psychoacoustics principles for this.

[TacoD]

This was six or seven years ago, and I can dig up the work if you want. It was fairly successful, but we only used ~10th order polynomials to approximate the 'ideal' transfer function for each driver. Matlab got unreliable in doing higher-order systems, probably something do to with precision problems with the built in algorithm.

Newer versions of Matlab have some additional handy functions in the signal processing toolboxes. Especially the algorithms to fit a FIR filter to an arbitrary magnitude or phase target are useful. If you export the coefficients and use these on a linux machine with brufefir convolution running, you are in business.

In the first trials I used a transfer function of a passive filter design from LspCAD in order to try to mimick a passive filter, however this was somewhat harder than a plug and play ultra-steep filter.

One big drawback in this methodology is that you need a copy of Matlab with the signal processing toolbox (makes life more easy), this is rather expensive (~1500 + 1500 euro) if you are not a student anymore or not working for a company with a license.