I came across a website where you can record or upload an audio file, and you get an automatically-edited version back, based on which of the several presets you select - an old-timey radio effect, a kidnapper-style voice change, a Dalek voice, etc.

Assuming that the modifications are not too outlandish or unpredictable (band pass filters, adding echoes/reverb, adjusting tempo/pitch, adding in particular sounds like bleeps or buzzing noises...), does there exist some sort of library of 'test sounds' which can be used to reverse-engineer the process and determine the parameters used to modify the sound files?


This is actually quite trivial. If you were to upload a "Sweep Tone" to the site, it will return the "sweep tone" modified for you.

The great thing about "sweep tones" is that they can be used to "reverse engineer" an impulse response which can then be used in a convolution filter to re-create the original effect.

Here is a link to an article I wrote which outlines the process.


| improve this answer | |
  • Useful answer. Am I wrong in assuming that your proposal is only valid for Linear Time Invariant processing ? And that many audio effects are not LTI (distortion, dynamics processing, modulation, ...). – audionuma Nov 11 '18 at 11:12
  • Although convolution is primarily associated with LTI, I am not 100% convinced all the examples you have quoted are excluded, but further research would be required in order to confirm that. Inverse convolution is a very common technique for analysing a signal path so I would assume this would get you quite a long way down the path in order to reverse engineer a processing module. – Mark Nov 11 '18 at 11:36
  • 1
    De-convolution is only effective for linear transformations, as convolution itself is a linear transformation and thus cannot be used to describe non-linear ones. While various filters, equalizers, delays and reverbs are linear, dynamics processors — compressors, limiters etc — are not. – bipll Apr 14 '19 at 1:43

I'm going to stick my neck out here & say, "Computers aren't smart enough to do that yet."

AI has come a long way, but I'm not sure it's yet smart enough to listen to a recording of a voice & say, "That's had a little soft-knee compression & possibly a pull in the low mids & a push at 4k"

At present I'd say there's no substitute for human experience, having heard all the components of those sounds before & using that as a first guess followed by some educated trial & error to recreate them.

Interestingly, all your examples would probably involve differing amounts of distortion, plus...

Old Radio - EQ changes; usually thinning the lows & pushing the mid/highs. [Crackles/phase-shifts optional]
Kidnapper - pitch-shift would be the Hollywood go-to. [Tinny speaker EQ & crackles for the usual phone call effect too.]
Dalek - 'pure' distortion, but I think it also has a touch of ring-mod in it to give the 'constant tone' feel.

| improve this answer | |
  • Mostly, what AI does is applying knowledge based on humongous datasets(in various ways). If lets say, a computer was loaded with all the possible sounds connected with a process that was used to produce them, it would be able to propose a set of processes and the probability of them being used to produce that sound. The more sophisticated it gets the more complex the sounds it can actually determine. Thats on a theoretical basis ofc.. Right now computers dominate the fields that need experience/static knowledge/or knowledge built over time (e.g. games - chess / alpha go etc..) – frcake Sep 28 '18 at 12:37

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.