A different speaker testing method

[Few]

Has anyone tried implementing the log sweep speaker testing protocol described by Farina in this paper? I'm not clear about whether it's an option in any of the commercial packages.

In a nutshell, a five second long exponential sweep of frequency (but called a log sweep) is used as the speaker stimulus; it's an exponential chirp. Then a time reversed and exponentially weighted version of the stimulus is convolved with the speaker's response. The result of that convolution is the speaker's impulse response, but in a novel form. Instead of a single impulse a series of impulses is generated. The latest one reflects the speaker's linear response to the stimulus. The earlier pulses correspond to the nonlinear responses--the nonlinear distortion. By inverse fourier transforming each of the separate impulses you can look at the linear portion of the frequency response, untainted by harmonic distortion, or you can look at any order(s) of the harmonic distortion. One of the claims is that the more common MLS method generates frequency response measurements that can be corrupted by nonlinear distortion, while Farina's method is less prone to that problem. If my description is too garbled to make sense, take a look at Farina's paper. It contains a more thorough and careful description.

I decided to try my hand at creating a speaker testing program that uses Farina's approach. As of late last night I managed to get a LabVIEW-based program working, and the whole idea seems pretty slick. However, since I haven't yet had time to apply it while actually designing a speaker I wondered if there are down sides that I haven't thought of or encountered yet. Does anyone else have any experience with Farina's technique? Is there a good reason it hasn't taken the speaker testing world by storm?

Few

[dwk123]

I don't know of any commercial packages that implement it directly, but there are several DIY versions that have floated around. I think Ed Wildgoose has a semi-portable version using either rtaudio or portaudio on his site (duffroomcorrection.com).

I had a partial implementation using BruteFIR, although I never quite got it automated - I had to manually select out the impulse from the resultant de-convolution, and didn't automatically compute distortion.

I agree though - very slick, and I'm surprised that it's not used more.

[houstonian]

http://gaydenko.com/qloud/

I've tried it, it's a nice package.

[Few]

Thanks, I hadn't seen the approaches either of you mentioned. Very interesting. I'll probably continue on what seems to be becoming a quest to create DIY speaker testing software for DIY speaker design (where does it end??) but it's reassuring to know others have found this path worth pursuing.

I take it neither of you (neither dwk123 nor houstonian) have found some fundamental flaw that suggests I shouldn't waste my time with the Farina approach?

My next step is to automate the process of calculating the SPLs associated with the first few harmonic distortion products. I was encouraged to find that the impulse responses for each of the harmonics show up right where Farina's work predicted. I'm (naively?) hopeful that that will make the automation process easier than it would be otherwise. It'd be pretty cool to have a series of harmonic distortion measurements automatically generated every time I do what otherwise would have been a simple frequency response measurement. Anybody know a way also to extract a single coefficient that expresses how good the speaker will sound? That sure would simplify the process.

[anli]

Originally posted by Few Anybody know a way also to extract a single coefficient that expresses how good the speaker will sound?

I'm not sure such coefficient exists at all.

BTW, which platform (Linux, win32, other) will be used in your program? The thing is, 'qloud' has the only Linux-specific class, if I remember (as author) well, and your can try to migrate the app to win32, as all other things are cross-platform (Linux/win32).

As for commercial apps, ARTA (win32 app) can use Farina's method.

[Few]

Yeah, the "good sound coefficient" was a weak attempt at a joke. It's just as well that no such thing exists--it would take all the fun out of it.

I intend to continue writing my own code, so my question really is more along the lines of whether others are using Farina's method, and whether there are any troubles with it that I should watch out for. My impression so far is that there are a few implementations out there, but it hasn't really become a standard method. If that's true, and the method has advantages over MLS-based systems, I'm curious about why the MLS systems are still the standard.

One feature of the log sweep I like is that it's possible to define the log sweep to start at any frequency so you're not forced to run low frequencies through a tweeter to test it. Obviously you can put a filter in series with the tweeter, but limiting the bandwidth of the stimulus seems like a more elegant solution. Plus, the idea that the various orders of harmonic distortion are neatly separated in time is just cool.

Few

[454Casull]

The GedLee metric seems to be fairly accurate.

[Few]

That's what I've read, but doesn't it really only address harmonic distortion? I'm asking this without access to any resources to double-check my memory, so I could be just plain wrong. In any case, I don't think that other things that affect the quality of reproduction, things like the flatness of frequency response and the directivity, are addressed. In that sense there still isn't a single number to express how well a speaker reproduces sound. Of course all this discussion is just a result of my attempt to make an audio joke. Serves me right.

Thanks for reminding me that I do need to go back and read up on the GedLee work, though.

[bwaslo]

Listen Inc.'s "SoundCheck" and Liberty Instruments' "PRAXIS" both implement the Farina method. (I'm the programmer for PRAXIS, but I'll try to keep this from being an advertisement!).

It is a very powerful technique. Farina's paper had most to do with measuring harmonic distortion, which the technique does very well and really quickly. Another benefit (and more important to me) is that it allows quick measurement of the frequency response while rejecting harmonic distortion products. It is also more immune than MLS methods to time shift issues (such as air currents when measuring in a large space).

If you measure something that has significant distortion (such as, say, a loudspeaker playing at moderately high levels) with MLS, the recovered impulse response portrays harmonics as weird fake pulses that look randomly interspersed around the impulse response, which can't be separated from the real impulse response, and so can mess up a frequency response test. With the Farina technique, the impulse response shows all the harmonics in negative time (i.e., before the main impulse) where they can be easily removed without at all disturbing the main "linear" impulse response.

If you ever need to get a clean loudspeaker impulse response (with very low corruption from distortion) the Farina method (also called "Log-Swept" or "Log Chirp") is the way to go. There is a technique in my software that can use a clean impulse response of a speaker to calculate it's linear effects (i.e., only frequency response) on a pieced of music. Then, the music is actually played through the speaker and the calculated linear recording is subtracted from the actual recording -- leaving a recording of only the non-linear distortion!

There is also an unconventional dual Log-Swept way to measure intermodulation distortions using a variation on the Farina technique, though it's pretty complicated and not much used.

>One feature of the log sweep I like is that it's possible to define the log sweep to start at any frequency..

Actually, you can't do that with the Farina technique as described above - you need to be able to calculate an impulse response from the sweep, which you can't do without sweeping the entire band (DC to half the sample rate). You can make a log sweep that starts at higher frequencies, but you can't calculate the impulse response from it.

A better way to avoid hitting the tweeter with high energy is to just drive it through a small series capacitor, which cuts out the highs. You can easily (and very accurately) compensate for the frequency response effects in the measured result by also getting the "reference voltage sample" from the far side of the capacitor. The rolloff drops out in the calculation and everthing stays happy.

[anli]

Originally posted by bwaslo
Actually, you can't do that with the Farina technique as described above - you need to be able to calculate an impulse response from the sweep, which you can't do without sweeping the entire band (DC to half the sample rate).

Why not? I do it in 'qloud'.