Drum Audio to Midi Conversion issues.

[Nunstummy]

I have a full drum track from an 80’s cover tune, but when I convert to midi I get many outlying midi notes.  The results are a mess and cleaning it up is more work than programming the beat in pattern regions.  Are there techniques or settings I’m unaware of?

I’m trying to capture the performance of the drummer, and convert to midi, so I can pick the samples for each kit piece.  If I program the beat in pattern region sections (verse, chorus, bridge, etc.) with Step Editor, I lose the performance elements of the drummer, like snare hits ahead or behind the beat, or hi-hat polyrhythms, etc.  I’m trying to maintain the style of the performance.

The isolated drum track is a result of an AI app that took an mp3 and broke out the parts into drums, bass, melodic instruments and vocals.  It did a reasonable job, but I cleaned up the drum track by increasing the gain to emphasize the transients and noise reduction to eliminate any artifacts.  So, I’m curious if I’m pushing the limits of Logic’s midi conversion of the flexpitch data.

A lot of the resulting midi notes fall below the default drum note assignments, and a lot are high-up, probably artifacts?  Logic’s midi conversion doesn’t know the source is a drum track, I presume it just finds ‘pitches’ that are different.  Pitch would be easy to detect when converting a melodic audio track, but the subtle different in pitch of real acoustic drums seems to be problematic. I noticed the kick pattern is spread between 3 midi note pitches and once identified, I can drag all those notes to the right kit piece, but after the kick and snare it gets really confusing, almost impossible to decipher.

Has anybody else had success converting an audio drum track to midi?  Logic or another 3rd party program?

 

[David Nahmani]

7 minutes ago, Nunstummy said:

when I convert to midi

How do you convert to MIDI ?

Have you tried Track > Replace or Double Drum Track? It lets you adjust the threshold. 

[Nunstummy]

13 minutes ago, David Nahmani said:

Have you tried Track > Replace or Double Drum Track? It lets you adjust the threshold.

Funny.  I’ve used that to double the kick or snare, but it never occurred to me to use it one kit piece at a time.  I’ll try.

[des99]

Software doesn't generally tend to do miracle work on tasks like this. Sometimes I find I need to faff around to manipulate things to convince the software to give useful results.

If your composite (mixed) drum track isn't being interpreted correctly, I'd probably try to isolate - eg, create multiple copies of the drum track, and gate out the core drums for each - one track for kick, one for snare, one for hihat etc. EQ them out if necessary (eg low pass the kick track, hipass the hats track etc), and use gates, strip silence and whatever editing tools necessary to provide as clean tracks as possible, bounce them down, and then pull the transients out of those using Logic's drum replace or doubling tools, and seeing how that works.

The more "help" you can give the automatic tools on the things they are struggling with, the better and more usable the results. I'd rather do that work on the way in and generate better results, than generate a bunch of poorly-interpreted garbage and then spend hours trying to clean that up afterwards...

Edited January 27, 2023 by des99

[Nunstummy]

41 minutes ago, Nunstummy said:

Track > Replace or Double Drum Track?

Not much better.  Even with threshold adjustments, the best I can do is kick and snare.  I've decided it's the source audio file quality that's generating the additional pitches that Logic interprets as midi notes.  The source has a lot of cymbals that ring through and overlap the kick, snare and other drums.  Setting the flexpitch to 'Rhythmic' shows how many extra notes Logic sees.

So, I'm trying to reduce the cymbals with EQ, apply the Waves NS1, and Logic Noise Gate.

But, I suspect Logic uses the source audio per-effects for the flex-pitch analysis.

I might try bouncing the audio file to be sure my efforts to reduce the noise are in play.

 

[David Nahmani]

1 minute ago, Nunstummy said:

Setting the flexpitch to 'Rhythmic' shows how many extra notes Logic sees.

Have you tried clicking the "-" (minus) button in the menu bar of the File editor so that you see less transients?

[Nunstummy]

12 minutes ago, David Nahmani said:

Have you tried clicking the "-" (minus) button in the menu bar of the File editor so that you see less transients?

Yes.  There's no quick miracle here.  Logic's transient analysis and flex-pitch are mistaking the snare for the kick or vica-versa.  They sound different to me, but not to Logic.  The best I can do is get the kick drum accurate by reducing the transients (- minus) and all the EQ, gate and noise reduction efforts.  Which means, I'd be creating the snare, hi-hat, toms and cymbals in step editor.  The snare is easy, the toms and cymbals are not too bad either, but the hi-hat rhythms will be challenging.  I'm certainly learning a lot about the limits of converting audio to midi - which is fun.

Here are 2 of the snip-its of the source files. The first is from the AI app isolation effort.  The second, after my noise reduction and EQ efforts. Still pretty noisy. These are MP3 due to the upload size, whereas I'm working with slightly better *.wav files.

 

 

 

 

[Nunstummy]

2 hours ago, des99 said:

I'd rather do that work on the way in and generate better results, than generate a bunch of poorly-interpreted garbage and then spend hours trying to clean that up afterwards...

Good advice!

[Nunstummy]

On 1/26/2023 at 11:44 PM, Nunstummy said:

I'd rather do that work on the way in and generate better results

Audio to Midi conversion works ok for a single drum track, or a monophonic melody.  Apple doesn't promise anything else.  So converting a complete drum beat, with multiple drums and multiple pitches is pushing the limits.  In my latest attempt, I tried to isolate each drum kit piece with EQ and the noise gate, hi/low cut-off frequency settings until all you can hear is the snare (for example).  Thinking, I would bounce-in-place individual tracks for each drum, then convert each to midi.  The approach has merit, but it really depends on the source.  In my example, the frequency of the kick and snare overlap.

Question: Is there a better tool in Logic to limit the frequencies than EQ or noise gate?  I'm looking for ideas?

[des99]

7 minutes ago, Nunstummy said:

In my latest attempt, I tried to isolate each drum kit piece with EQ and the noise gate, hi/low cut-off frequency settings until all you can hear is the snare (for example). 

Yes, this is exactly what my advice to try was. Sometimes you need to try a combination of techniques to get what you need, rather than expect a one-button miracle 🙂

[Nunstummy]

Saga continued:  Somebody recommended RipX by Hit'n'Mix software.  I just downloaded the free trial.  Imported the same mp3.  The drum isolation is pretty good, with 3 layers - Kick, Drums, Percussion.  It also has "Export to midi" - a more direct route than Logic's Flexpitch + Convert Pitch Data to Midi.  The midi file is fairly clean, but has wrong notes and a messed up tempo.  I like that it only generated midi notes for 4 or 5 pitches, but it mis-reads the transients so the pattern/rhythms are incorrect.... but maybe this could work...

I also dragged in the drum only *.wav file it created, then applied the usual Logic FlexPitch analysis and Convert to Midi.

Once again, it creates too many midi notes, deciphering many different pitches, but it's the cleanest so far.

Every time I use Logic "Convert to midi from pitch data":

1. The velocities are always "1" so I have to select all, and set the velocities to 100 or something I can hear.

2. The midi notes are 1 octave too low for a standard drum instrument, so I select all and move +12 semis.

This would all be much easier, if every kit piece could be isolated.  Most of the AI tools separate drums from vocals, bass and other.  I need AI that separates a drum beat track or loop, into individual kit pieces.  Not sure that exists.  The challenge is that drum pitches can be close together.  A kick can sound like a floor tom, a snare can sound like a small tom, hi hat open can sound like a crash or ride - it's all too tight.  An AI model would have to predict what kit piece to sound based on some advanced logic, like hi hat it often an 8th or 16th pattern, whereas a crash is at the end of a phrase.

Still no "Miracle" yet.