Maddest Sanity - Restorative Reconstruction

What’s so special about this particular album?

This is the most important album in my life and audio engineering journey. It led to me returning to music after Joy of Creation (my childhood music), then starting Number Eleven, Electron Emitter and then worked as a confirmation that I can open Electron Transition.

Why? It’s the album I’ve listened to since childhood, because it was composed by my older brother – eww. This album has shaped my music understanding as a language from the earliest days.

When I was considering my life choices in the reverberant corner of my bedroom studio, I figured that I need to choose things that would lead to one of these two things:

I have enough moneys to hire someone to mastergoodize it so it would sound as good as it could’ve been.
I have enough experience to do that myself.

It led into the second option.

Before and after is at the bottom of this page.

Why did it need Restorative Reconstruction?

There are only two pictures of those on google. I only have one speaker, making a mold for the second one because I’m mad.

First of all – it was composed with these, so it sounded shit.

Second of some – it was distorted, because for some reason both of us didn’t have the knowledge why clipping is bad. And it was clipped a lot.

Third of and then some – it was exported as low bit-rate mp3 files.

Fourth of them and even more – there was no initial sound design or mixing done to them, because we didn’t know anything about engineering or sound at all.

I knew it sounded dry, bleak, weak, boring, flat and other synonyms for terrible.

Why is it ridiculous?

Stem separation wasn’t an unheard concept for me, but I knew that’s not going to be enough, because individual drum samples that were used, were out of balance. In order to re-balance the drums alone, I’d need to recompose them and then mix in reference with the track. But I wouldn’t be able to re-create the album exactly by myself.

I wouldn’t be able to extract them clean as well, because they were buried under lossy compression, distortion and then everything else – synths, bass and sometimes effects.

Stem separation would be separating the cake into distinct layers from which it consists of.
This, would be taking the eggs out of the cake, after it’s been baked.

And here I am, questioning can I take this challenge, when I didn’t exactly knew how to do it – make this album sound as good as it could’ve been.

The deconstruction – how did I separate the entire album into individual samples:

1 – I converted the tracks back into 32-bit float wav files and de-clipped them.

Since the tracks were distorted because of clipping, I understood that I have missing information above 0 dBFS. Let’s approximate that with de-clip.

2 – Deconstruction.

2a – why and when it’s possible

This album was composed in Fruity Loops 3, meaning that it’s completely looped pattern based, didn’t use too many non-linear effects (like phasers, flangers, choruses), that wouldn’t sync with the rest of the grid, allowing me to heavily rely on simple phase cancellation.

It was also pretty progressive in its composition – you’d have dry drums first, then sometimes a single pad as a gap in the composition, things would appear on their own, rather than everything together in a drop. The way it was composed helped a lot to extract it clean.

2c – the first extraction

I had starting samples like a pad or drums to extract completely clean because they were alone – cut that out of the track, flip the phase and then subtract with phase cancellation from the rest of the track.

2d – the following extractions

The removal of samples by phase cancellation, leaves other samples alone. When possible, I repeated this process until I was left with things that are only appearing together, have non-linear effects, or when I encountered audible artifacting from either – lossy compression or distortion.

2e – cleaning up the extracted audio

When I encountered extraction artifacts – I cleaned the sounds using only linear phase EQ, spectral de-noise or deconstruct. Restored the samples back into their original waveforms as much as possible.

Note – lossy compression will leave a lot of artifacts in the high frequencies, depending on the bit-rate. If that’s the case – aim to remove the low frequency things first.

Progressive composition, instead of everything everywhere all at once.

A melody, extracted from a noisy leftover of drum ambience

Hear all the leftover artifacts and drum ambience.

2abcde – example:

The track starts with a pad, looping for four bars. Then the drums appear, then the bass.

I cut the pad out, invert the phase and place where it’s present in the track. I’m left with clean drums.

I cut the drums, individually if possible, then flip the phase, I’m left with the bass, but with artifacting from the non-linearity of lossy compression, but it’s only at the high end.

I cut the bass, apply linear phase filtering to remove the artifacts in the high end.

2f – Summing other cuts to extract

I ran out of individual samples. This is a bit more complex – I will refer to individual samples as ABCD.

You only have A alongside B. AB.
You only have B alongside C. BC.
You only have C alongside D. CD.

This situation is not limited to AB, BC or CD.

You can AB – BC = A-C.
You can A-C + CD = AD.

I made more options by summing and subtracting. The objective here is to get a result that’d give the easiest extraction.

Identify which one can be edited out the cleanest. This goes in order of the cleanest extraction:

You have the original samples and you can match the alterations done to it.
The samples are not overlapping in the frequency domain.
The samples do not have a lot of overlapping harmonics.
One of them is not a complex waveform and can be separated using spectral edits.
It can be deconstructed or AI’ed from. This works only if you have a noisy sample alongside a harmonious sample.
Chewing off your work desk’s lacquer and spending the rest of the day doing spectral edits.

Properties that make the editing out a lot time consuming and harder:

Both samples are of high fundamental frequency, or have a lot of harmonics in the high end.
Both samples are harmonious.
Both samples are non-linear or have a lot of FX on them that doesn’t loop or sync.
Both samples are noisy and overlap in the frequency domain. This will be impossible.

I labeled all the possible options, from ABCD, which one of those can be separated the easiest. I aimed for no overlap in the frequency domain, both being simple, or easily seperAI’table.

An example of 2f:

You identify that A and D would be the easiest to separate.

AB – BC = A-C
A-C + CD = AD
Separate A from D.

Now you can separate A from B and then C from B, with much cleaner results.

AB – A = B
BC – B = C

Clean every each of them of the separation artifacts.

2g – left to be the last

Some things will be the last ones to be extracted.

A pad that had non-linear effects, like a chorus on top, or a long, 32 bar melody, or something that never repeats, this will include all live instrument recordings, if not looped.

Everything will have to be subtracted away from it, in order to have it alone.

This required the most amount of work, because I was working through the entire length of that sample, rather than a single impact, 2 or 4 bar loop and it left a lot of editing artifacts.

Get it to be as clean as it can be.

2h – Hold on a second! What about overlapping decays, reverb or delays?

Ha! This is where I got to gnaw at my fingernails and got a pair of big, fat and wet black circles around my eyes. Because I was dealing with loads of leftovers, artifacting and possibly being unable to extract things clean.

Most of the things will never extract clean, will leave loads of leftovers and it will limit usable extractions to around 2-3 samples.

You can forget the 2f option for clean extraction, to sum or cancel things that are together, unless they feed into the same reverb.

If the reverb has any sort of modulation or randomness – forget it completely. De-reverb will not go that far and if you have delays – it will be an actual motherfucker to deal with. You won’t even get clean samples themselves, because they will layer with the delayed things from before.

This is where everything that I know about restorative practices and all of my skills to resist the intrusive thoughts to drop everything, while tediously doing spectral edits came into play.

The leftover you’d have from drums.

Leftovers with some extraction artifacts.

Bits of everything left over. (this is a very solid extract though)

Lots of mp3 and extraction artifacts.

Just cry at this point.

2abcdefgh notes:

Try and achieve the cleanest possible extract. The most important part of the sample is the transient, then the decay and only then the ambience.

Every sum and cancel will raise the noise floor, depending on the quality of the track (if it’s 32-bit float, you can ignore).

Every sum and cancel will increase the likelihood of artifacts, particularly if it’s a lossy or distorted track.

You may have leftover bits that were overlapping when separating things. You can use that separately as an overlapping layer, leave it to use for layering to mask the effects of separation.

The reconstruction – how I put everything back together:

3. Planning the objective – how to reconstruct

Now when I have everything separated as much as possible, I subtracted that away from the master, the initial track. You should only hear the leftover artifacts. If it isn’t perfect and you have remainders of the reverbs, delays, or some decays, that’s not a big issue.

You may even want to clean the remainders, to use them later on.

Now – the approach on the reconstruction will differ, depending on the material and why it’s being done.

If the results are perfectly clean, great, it can be mixed like any other track.

If there is a lot of remainders and the reconstructed track is audibly reconstructed – there will be a lot of masking and purposeful sound design decision making, which might change the character of the track.

The difficulty in reconstructing Maddest Sanity came from the fact that I’ve been listening to this album for more than 20 years already, which made me terribly biased towards it’s imperfections.

And I didn’t want to change its character in any way, the objective was to make the tracks sound as good as they could’ve been, not “goodize” them up to some standard.

The tracks themselves didn’t have too much effects or “ambience” that would give a nice musical noise floor that could bury all the artifacts. They were completely flat and empty, digital and simple. Any attempt to add any kind of “ambience” to mask artifacts would change the tracks. There was a lot of silence in between the samples.

Therefore, I had to come up with a sound design idea for the album, that didn’t have one in the first place and it had to be transparent, in order to not change the tracks.

This would’ve been my reaction in 2005 if someone would’ve told me I’d be doing this reconstruction.

I ended up making three distinctive sound design ideas for their spaciality:

Simple panning.
Haas effect that’s not 100% wet.
Created with reverb.

The root idea being – primitive. Depth is created with just EQ.

These ideas came from the tracks themselves when grouping them – reverby, localized or dry.

4 – The reconstruction

Kick routing – ignore the gain levels on sidechain knobs. I control that inside Patcher.
And I’m not using LAYER on Patcher.

Advanceder processing for a very low quality snare.

Imagine this routing, for 30 individual samples, using the mixer, instead of Patcher.

You’d run out of mixer very quickly.

Since the album would not deconstruct easily and would leave a lot of lossy compression, or de-clip artifacts, leftover decays, remainders from overlaps, I opted to go with the layering approach.

4a – I set up a parallel processing patcher, with a very specific routing.

I had five main channels, or more, to separate drums for instance.

INIT – the initial track.
CANCEL – flipped polarity – I route things I want to subtract from the INIT.
ADD – things I add to the INIT.
SUM – I sum the ADD, CANCEL and INIT on this channel. There I can do some processing.
LAYER – effect layers, like transients, or bit-crushing.

4b – Rebalancing the mix

And the way the mixing went, was the least intuitive way I ever had. I had my patcher set up, where the output of an extracted sample is routed into CANCEL, ADD or LAYER.

Every single process had to be linear phase, in order to nicely add or cancel with the INIT.

I had to preserve the weak high frequencies of the INIT as much as possible, so I had to arrive at a balance, by subtraction in the low-mid frequencies by gaining them in the CANCEL path of the patcher.

If I wanted to boost something, I could boost that in the ADD channel, or boost everything else around it in the CANCEL channel.

In the ADD channel I usually added stereo information – I sent things that were either panned hard left, or had the Haas effect, until I reached the desired position in the stereo field.

4c – Creating fidelity and detail

On the LAYER channel, I’d do heavy transient processing, bit-crushing, re-aliasing, distorting and every other way I knew, to bring back life into the high frequencies, that got destroyed by the lossy compression.

If the track had delays and reverbs, there was no way for me to extract them. I did however knew the exact effects that were used and I matched the settings as much as I could, to add more of said effect in the frequencies, or stereo field as much as I needed to.

I also managed to locate some samples that were originally used. I used them to layer on top of the extracted ones, sometimes passing them through the same effects, to layer with the INIT, to mask the lost fidelity due to artifacts.

I was unable to arrive at a balance I wanted immediately, due to insane bias that I had for the imperfections of the album. What helped is that I could adjust, or do additional processing on the buses.

4d – Removing unwanted leftover artifacts.

When I was happy with the mixes and the way the tracks sounded on every single toaster, I proceeded to shove them into Isotope RX and listen, after drinking several bottles of energy drinks and eating the bottles, so I’d hear every single click and pop, woop and pre-ring that happened due to lossy compression.

So, some hyper-attentive listening and lots of spectral editing.

4e – Mastering.

Since I spent a couple of months working on something I listened to for the last 20 years, I decided to take a lengthy rest from this project and I mixed I Follow and The Cat Has Cometh! Came back with reset ears, did a very easy master and then I compared the initial versions with the masters.

Before and after is available on the bottom of this page.

This project served as an exam for myself, because I thought that I’ll never be able to do anything like it. Why I was able to do it though?

massive breath in preceding a cheesy motivational quote

I needed more transients.