MIDI files from Audio using AI

190 views
Skip to first unread message

spencer chase

unread,
Mar 23, 2024, 2:27:52 AM3/23/24
to Disklavier Support Group
I have spent 9 months writing and perfecting a program to improve the output of an AI program that converts solo Piano performances to MIDI. The raw output of the AI program is quite good but lacking in certain details. My program fixes many of the deficiencies and the results are quite amazing. I am offering a set of 17000 MIDI files made from my collection of CDs obtained over 20 plus years of collecting as well as music from the makers of the AI program which have been "improved" with my program. You can find information, listing of titles and samples on my web site at the following.

Sheldon H. Deluty, M.D.

unread,
Mar 23, 2024, 7:11:37 PM3/23/24
to 'spencer chase' via Disklavier Support Group

good morning:

can you explain in simple terms how you actually do this?

can this program in theory listen to any piano music and convert it into midi?

and what does it do when it hears other instruments at the same time? or can it only do your magic when the source is solely piano music

are the cd's or lp's played "over the air" and a microphone picks up the sound and then converts that into midi?

i apologize in advance if my questions seem dumb or obtuse...i have owned a Mark II disklavier since 1991 and have taken it through software and hardware upgrades over the years and I am absolutely fascinated by what you are describing...i would just (for my own edification) like to understand more about the technical aspects of this brilliant and groundbreaking achievement you have made

i have tried programs that scan sheet music and convert the graphic images into musical notation and then into midi but the results have been very disappointing

from the little i understand of this process, i am astonished at your achievement

thanks for your time

sheldon deluty



On 3/23/2024 2:27 AM, 'spencer chase' via Disklavier Support Group wrote:
I have spent 9 months writing and perfecting a program to improve the output of an AI program that converts solo Piano performances to MIDI. The raw output of the AI program is quite good but lacking in certain details. My program fixes many of the deficiencies and the results are quite amazing. I am offering a set of 17000 MIDI files made from my collection of CDs obtained over 20 plus years of collecting as well as music from the makers of the AI program which have been "improved" with my program. You can find information, listing of titles and samples on my web site at the following.
--
To seek membership to this group, send an email to disklavier-suppor...@googlegroups.com.
---
You received this message because you are subscribed to the Google Groups "Disklavier Support Group" group.
To unsubscribe from this group and stop receiving emails from it, send an email to disklavier-support...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/disklavier-support-group/51602680-4a9d-4a1f-9449-a3ef1c56f355n%40googlegroups.com.

Geoff Ward

unread,
Mar 24, 2024, 4:08:01 AM3/24/24
to Disklavier Support Group
Sheldon, I will answer your queries as I helped Spencer develop his software.  You start by creating a midi file from an audio file using software developed using AI by 4 Chinese students at a Chinese university.  I will include a link later.  I usually capture the audio from a Youtube video of a performance I like.  Youtube audio works pretty well.
The Audio to Midi (A2M) software is claimed to work best if the piano was a Steinway and the recording quality is very good.  However, quite surprising results can be obtained from not-so-good recordings of other pianos.  For instance, I got quite good results from a couple of pianists I like playing old uprights which were almost honky tonk.  If there are other instruments, including drums, included with the piano, the software will attempt to convert them to midi piano notes and the results are usually poor.  This where an online service called FADR.com is useful.  FADR can split an audio file into separate audio files (called stems) of piano, vocal, drums, bass, strings, etc.  It works pretty well, but it costs $10.00 USD a month.  There is another service offering from moises.ai which is cheaper than FADR, but I had a dispute with them over the charge they hit me with and now will not use them.  Their offering is not as good a FADR.

OK, so you have a piano audio file and you want a midi file.  The A2M software is very simple to use - no parameters to set - just output folder and source file.  On my computer, it takes a long while to open - about 2 minutes.  It takes about double the playing time to create the midi file.  A whole folder of files can be processed as a batch.  If it is a complex classical piece (see this one - the A2M results were very good - https://www.youtube.com/watch?v=LdH1hSWGFGU&t=1s           ) you will probably notice a couple of things.  1.  Fast repeating notes are very close together and some solenoid pianos won't play them properly.  2. The software has trouble distinguishing sustain pedal actions from notes held down by the pianist and will not release the sustain pedal with enough time to dampen the strings.  This is where Spencer's software comes in.  It attempts to fix these issues and produces a file that sounds much better.

I think that answers your questions.  You can download the A2M software and Spencer's programs from https://www.spencerserolls.com/A2M2.html

Kind regards

Geoff

Sheldon H. Deluty, M.D.

unread,
Mar 24, 2024, 6:30:19 AM3/24/24
to 'Geoff Ward' via Disklavier Support Group

thanks

so there is no windows desktop version?

sheldon

Spencer Chase

unread,
Mar 24, 2024, 1:41:41 PM3/24/24
to disklavier-s...@googlegroups.com
what? the A2M software is for the windows desktop as well as my "improve" program.


------ Original Message ------
From "Sheldon H. Deluty, M.D." <md...@earthlink.net>
To "'Geoff Ward' via Disklavier Support Group" <disklavier-s...@googlegroups.com>
Date 2024-03-24 3:30:15 AM
Subject Re: MIDI files from Audio using AI
--
To seek membership to this group, send an email to disklavier-suppor...@googlegroups.com.
---
You received this message because you are subscribed to a topic in the Google Groups "Disklavier Support Group" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/disklavier-support-group/kY4uN9LpHY8/unsubscribe.
To unsubscribe from this group and all its topics, send an email to disklavier-support...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/disklavier-support-group/3e14c0a7-b996-47cb-a574-b79a4523d591%40earthlink.net.

Sheldon H. Deluty, M.D.

unread,
Mar 24, 2024, 1:47:47 PM3/24/24
to 'Spencer Chase' via Disklavier Support Group

is this not the website for the software?

https://beatbars.com/en/a2m-au.html

do not see a windows option

or am i at the wrong webpage?

sheldon


On 3/24/2024 1:41 PM, 'Spencer Chase' via Disklavier Support Group wrote:
A2M software

Spencer Chase

unread,
Mar 24, 2024, 2:00:09 PM3/24/24
to disklavier-s...@googlegroups.com
I don't think this has anything to do with what Geoff might have linked. I also doubt it could process complex piano as does what I have been using. If it does and if it does not have (very unlikely) the anomalies that the software I used then the work that Geoff and I did was a waste. I suspected that some day better software that did not require the output to be edited would be available but I seriously doubt that what you linked is it.


------ Original Message ------
From "Sheldon H. Deluty, M.D." <md...@earthlink.net>
To "'Spencer Chase' via Disklavier Support Group" <disklavier-s...@googlegroups.com>
Date 2024-03-24 10:47:44 AM
Subject Re: MIDI files from Audio using AI
--
To seek membership to this group, send an email to disklavier-suppor...@googlegroups.com.
---
You received this message because you are subscribed to a topic in the Google Groups "Disklavier Support Group" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/disklavier-support-group/kY4uN9LpHY8/unsubscribe.
To unsubscribe from this group and all its topics, send an email to disklavier-support...@googlegroups.com.

Sheldon H. Deluty, M.D.

unread,
Mar 24, 2024, 2:01:15 PM3/24/24
to 'Spencer Chase' via Disklavier Support Group

so where can i find your program, please?

thanks

sheldon

You received this message because you are subscribed to the Google Groups "Disklavier Support Group" group.
To unsubscribe from this group and stop receiving emails from it, send an email to disklavier-support...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/disklavier-support-group/em371a424f-a06d-4292-bc55-ab48371642a6%40390cf305.com.

Spencer Chase

unread,
Mar 24, 2024, 2:02:25 PM3/24/24
to disklavier-s...@googlegroups.com
I looked at the program you link and at the end of the page is says "no polyphony" this is a simple pitch recognizer nothing more and will do nothing with complex piano music.


------ Original Message ------
From "Sheldon H. Deluty, M.D." <md...@earthlink.net>
To "'Spencer Chase' via Disklavier Support Group" <disklavier-s...@googlegroups.com>
Date 2024-03-24 10:47:44 AM
Subject Re: MIDI files from Audio using AI
--
To seek membership to this group, send an email to disklavier-suppor...@googlegroups.com.
---
You received this message because you are subscribed to the Google Groups "Disklavier Support Group" group.
To unsubscribe from this group and stop receiving emails from it, send an email to disklavier-support...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/disklavier-support-group/695c000a-9ad1-4f11-bcb2-d854f1fc2632%40earthlink.net.

Spencer Chase

unread,
Mar 24, 2024, 2:03:02 PM3/24/24
to disklavier-s...@googlegroups.com
------ Original Message ------
From "Sheldon H. Deluty, M.D." <md...@earthlink.net>
To "'Spencer Chase' via Disklavier Support Group" <disklavier-s...@googlegroups.com>
Date 2024-03-24 11:01:13 AM

Sheldon H. Deluty, M.D.

unread,
Mar 24, 2024, 3:07:46 PM3/24/24
to 'Spencer Chase' via Disklavier Support Group

thank you

sorry for the misunderstanding

Spencer Chase

unread,
Mar 24, 2024, 10:24:08 PM3/24/24
to disklavier-s...@googlegroups.com
Have you listened to the samples to see how good they are? They really are good. 

this page has some stuff you can download including the programs that made the files. I think there are some descriptive docs about the process and who wrote the programs etc.

i did not write the AI program. that was written by some Chinese students as a dissertation i think. Then a friend of a friend took that program and made it usable without being a total nurd. however the results as good as they were are far from perfect. overpedaling missed repeats stray notes ineffective pedal codes and so on. also there were unnaturally extended notes while the pedal was on that did not in and of themselves change the music but made it very difficult to edit. and also the excessive polyphony was a problem for player pianos. find with a synth but many players limit the polyphony and or may do nasty things like blow fuses.

together with Geoff Ward as a tester I wrote a program we call A2M Improve which fixes these problems very well. 

i bought a high speed gaming computer for a couple thousand dollars and let it loose on my collection of audio piano solo CD rips that i have collected over more than 20 years. the AI program only does solo piano but there are programs that work pretty well to sort out different instruments. in fact Geoff has had some success with one of the programs and did a great job of making a demo video in which the video is synchronized with the MIDI and other instruments were removed.

you can get the tools in the link below and do it yourself. With the help of Geoff it took about 9 months of programming and testing to get the program to work well enough to be pleased with the results.   using the default settings you can get similar results with most well recorded music. Some stuff does not work such as doctored music with added reverb or bad miking or added effects. 

all you have to do is collect hundreds of CDs of piano music or spend many hours streaming stuff online some of which works and some of which doesn't get a $2000 computer (or spend weeks waiting for the processing) figure out how to use the programs including mine (which does have a good help file) 

oh and also write programs to extract meta data from MP3 (and convert from other formats to MP3) and insert it back into the MIDI files. if you like titles (with no internal meta data) like CD33499-977AM you don;t have to bother with this and you can title and tag files all by hand.

i have been messing around with FLAC MP3 and MIDI files for 25 years and have amassed a huge collection of tools to make this all possible. that is why i can offer 17000 files for $165

i wish the meta data were better but I was an early adopter of CD ripping software and in the beginning the databases for meta data were really horrible. that is why i have several performances by Chopin and Beethoven. I do not have the time to go through the 17000 files and make corrections but If my health gets to the point that i am no longer mobile and can only sit at the computer all day i might consider making corrections to obvious thins like played by Chopin :)



------ Original Message ------
From "Sheldon H. Deluty, M.D." <md...@earthlink.net>
To "'spencer chase' via Disklavier Support Group" <disklavier-s...@googlegroups.com>
Date 2024-03-23 7:17:49 AM
Subject Re: MIDI files from Audio using AI
You received this message because you are subscribed to a topic in the Google Groups "Disklavier Support Group" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/disklavier-support-group/kY4uN9LpHY8/unsubscribe.
To unsubscribe from this group and all its topics, send an email to disklavier-support...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/disklavier-support-group/55d47053-c747-4be6-9e0e-0402574bd91a%40earthlink.net.
Reply all
Reply to author
Forward
0 new messages