Forensic frequency response

[Blumlein 88]

Okay, sorry to disappoint, but no CSI here exactly. No dead bodies.....yet.

Wanted: critique, commentary or explication of whether something I do makes sense or not.

By forensic I mean determining from musical recordings the relative frequency response of two or more pieces of gear. A couple of examples. Let me say I play the same musical source through two pre-amps and record the output with an ADC. Or perhaps the same digital file through two DACs and record with an ADC. Can you use the music to determine relative frequency response of the gear?

What I do is select a few seconds of the more energetic parts of the music, and run an FFT on it. Often a 4096 point FFT which gets me 2048 frequency bands. I make sure it is done on the same exact parts of two recorded files. I export the values of each band, dump them into a spreadsheet and subtract one from the other. You then get a curve of difference between two pieces of gear.

This seems to work to a fractional db level. Seems quite accurate, repeatable and reliable giving the same results with multiple pieces of music. Also thus far in the few instances I have checked gives the same answer as running a frequency sweep or spot tests of sine waves.

You do have to pay attention to details of level and noise. If the recording has no content at frequency extremes then you can't do a good comparison. You are mostly comparing noise levels so low the results make no sense and are of no use. Conversely the devices under test must be relatively low noise. Otherwise the same issue you are comparing noise rather than response to input signals. So quite often you can only compare over a 50 or 60 hz to 12 or 15 khz range. Still that is most of what we care about.

Also, though this isn't the motivation for me doing this in the first place, it occurred to me later it answers one common critical complaint of measurements. That the frequency response was determined with sine waves and not music. Well so far it seems dynamic music and sine waves or sine sweeps give the same answers.

So what say some of you with more knowledge, training or experience in measurements. Any pitfalls or problems I am not accounting for in this method of measurement?

 

[Opus111]

You say 'a common critical complaint' is that 'frequency response was determined with sine waves and not music' yet your instance of it is the first I've ever heard. Sure I've heard (and generated myself) plenty of complaints about testing distortion with sinewaves rather than music, but never FR. So do you have a link to an instance of this out in the wild?

 

[Blumlein 88]

You say 'a common critical complaint' is that 'frequency response was determined with sine waves and not music' yet your instance of it is the first I've ever heard. Sure I've heard (and generated myself) plenty of complaints about testing distortion with sinewaves rather than music, but never FR. So do you have a link to an instance of this out in the wild?

It is something I have read a number of times. In the end, who cares. What I am interested in is the method I outlined above. Yes, most such comments are about distortion.

 

[Opus111]

The standard method adopted by Audio Precision is a swept sinewave, however that's not particularly quick for production test so when speed is of the essence, a multitone signal is used with 30 or more discrete frequencies present.

 

[TBone]

By forensic I mean determining from musical recordings the relative frequency response of two or more pieces of gear. A couple of examples. Let me say I play the same musical source through two pre-amps and record the output with an ADC. Or perhaps the same digital file through two DACs and record with an ADC. Can you use the music to determine relative frequency response of the gear?

Yeah, this has been of interest to me lately; take two or more cd players with different "apparent" tonal characteristics. Rip each player playing identical music. Will the numbers verify / justify what's heard?

 

[Blumlein 88]

The standard method adopted by Audio Precision is a swept sinewave, however that's not particularly quick for production test so when speed is of the essence, a multitone signal is used with 30 or more discrete frequencies present.

Yes, I get that. Familiar with the multi-tone test for speed.

This isn't about speed though. It is about having files, and not the equipment around to test. That is why I called it forensic. The gear is gone and not here anymore.

When I first did it, I had identical source material that was run through several pre-amps. All, except two sounded identical to me. Wondering about the two that were different I tried this on the files. All those that sounded the same had response numbers within .2 db of each other and mostly within .05 db of each other. The two different sounding pre-amps showed a different response. One being - 6db apparently at 20 khz, and the other shelved a bit at 2khz and rolled a bit on the low end. Which also fit with how they sounded though I couldn't have described it with any precision.

The other place it might be useful though results are sometimes less clear is comparing different masterings of recordings. If they haven't been compressed differently you get results that seem to make sense. (admittedly an uncommon situation that newer masterings aren't compressed more).

 

[amirm]

From my quick read , seems like a reasonable approach. I just wonder about level and time matching of the two outputs. The latter with windowing of FFTs could show differences that may not be there.