Can you recommend a good MP3 to sheet music converter program that is accurate?

Could be worth a read:
https://windowsreport.com/automatically-transcribe-music/
But I think that humans probably do a better job than computer programs at present.

Have a look at this
Meet Basic Pitch: Spotify’s Open Source Audio-to-MIDI Converter
https://engineering.atspotify.com/2022/06/meet-basic-pitch/
https://basicpitch.spotify.com/

Frank,
I spent the better part of yesterday trying out many of the converter programs. Including Basic Pitch, Transcriber, and Amazing Midi. Basic Pitch produced the closest sounding midi. But the notation was unusable, Completely. Even though playback was not too bad. It seems to me that these programs are meant to produce a file that is used in a DAW, and can be manipulated there. Not in notation software.

For testing, I used a simple four part fugue that I wrote for piano. The wav. file was clean and without any effects.

Transcriber produces a midi but you can't save or download it.

Amazing midi hasn't been supported since 2003. There are no instructions on how to use it.

We are a very long ways away from this being viable AI technology. Check back in a decade or two. Right now, human ears are 10000000 times better. Myabe each year you can knock a zero off that...

Two ideas to try:

1. Try Frettable. Free version available, then subscription with longer length.
2. Run it through an equalizer first. This way you can reduce the clutter and pick out one instrument at a time.
3. If a song is available in multiple covers, doing the extraction from multiples may help.

Instruments are picked out by lookng at the harmonic structure. This gets really messy in an ensemble. Some instruments are a lot worse, I suspect.

• Electric guitar with distortion.
• Drums. Picking up the beat is easy, but picking which drum can be tricky.
• Instruments that play with pitch bend, wah-wah

As a workflow you may need to bring it into a DAW that allows you to edit a "piano roll" midi and then you can move notes from one instrument to another. Check your work by having the mp3 as a track, then play it on one ear, and your corrected midi in the other ear.

Given how difficult a task this is even for someone who is musical, I'm amazed it works at all.

https://pianoconvert.latouchemusicale.com

Hey,

You can try PianoConvert, which is currently the most accurate software for transcribing polyphonic solo piano audio into scores and MIDI files (works with YouTube link or audio file): https://pianoconvert.latouchemusicale.com/

I've heard that they're going to extend use to instruments other than the piano soon, but I don't know more than that

Hi !

I am NOT an expert in this subject, but I have been interested in this for many years, and I have purchased several softwares packages aimed to transcribe ( among them Akoff, AnthemScore) and evaluated dozens of others ( TSAudio_to_midi, IntelliScore, DigitalEar, Amazing midi,etc, etc).

In my opinion it will take a few more "decades" before we will be able to get a decent sheet score out of an mp3. Let me explain some of the challenges, as I see them:

• music notation is more "symbolic", more "abstract" than for example, piano roll. In notation You can write a note as F# or Gb but from a theoretical point of view one of those forms could be wrong ( you need to analyze the current tonal center, the function of the note -if it is appoggiatura, or a passing/in-between note, the rules for engraving, etc.) All these aspects are difficult/demanding to program.

• I have seen professors arguing over how to notate/interpret a chord, specially towards the end of the tonal period (List, Wagner, etc). For example the endless discussions about Wagner's "tristan chord". You can see in youtube a few videos about this discussion. And it is just a chord !!!

But OK, we could live with these limitations and go and manually edit the score correcting enharmonics and renaming chords.

But there are more challenges:

• Look at the attached picture. if you for example play in a guitar a chord and let it ring, what sounds does not match the standard way of writing it. This is an important concept: notation does not always show what sounds, but instead it is like a human-friendly instruction on how to produce the desired sound. To program that in a software will be very demanding.

• Detecting "reallentando" or "accelerando" and properly notating it is a challenge. I have not seen any package detecting that.

• Often human transcribers - even those with profound knowledge of music - disagree if something is to be written in for example 3/4 or 6/8..sometimes I see a lot of academic discussions about this (sometimes the rhythm goes in 3/4 but the harmony changes in 2/4...or two instruments are playing in 3/4 resp. 6/8).

• Many jazz, new age, flamenco musicians play long passages of meterless music (not to name most of arabic, persian, turkish music), which confuses both humans and computers. I doubt that musicxml format supports as of today meterless music (I might be wrong in this), and most software packages support the notation of meterless music by means of "workarounds" ( like creating gigantic and weird time signatures e.g. 247/8 and hiding them, etc. Maybe only dorico -out of the big software packages - has native support for meterless music). So there we have additional limitations in the lack of a general, open, standard notation format capable of notating meterless music.

• Similarly, a lot of music uses microtonality, "quarter tones", and once again I do not think there is currently support in musicxml for quarter tones music. Actually there is not a standard way of notating microtones either. In the book "Inside Arabic Music" you can see both older and more modern ways of notating microtones.

• if a recording contains several instruments, the frequency spectrum becomes cluttered and it is difficult - even for a well trained human, to separate notes from overtones, and which notes are coming from which instruments. There has been some progress in AI in the area of source separation, but it is not perfect yet and it sometimes fails.

• If a guitar in a recording is too much "lost in the background" because of other instruments sounding much louder then transcription becomes some kind of "reconstructing", guessing what you cannot hear, making educated guesses. Such is also difficult to program.

• Let me give you a concrete example: you know the song by Phil Collins "Against all odds"? The introduction is just 2 measures of Phil collins playing the piano with a DX7 in the background. The first, natural reaction would be "how difficult can it be to write that down?". Well, I have seen 4 different notations of the introduction. Transcriptors disagree in what Phil Collins is playing in the left hand (some believe he plays octaves, others believe that he plays single notes.) The transcribers also disagree as which notes are played by which hand. Etc,etc. If 4 professional human transcribers cannot agree.....what can we expect of a software ?

• Some things can theoretically be played in a guitar but not practically ( like...say.... stretching 7 frets). A violin played with a bow cannot play 4 notes at the same time,etc. All these knowledge needs to be programmed as well. Each instrument has its own limitations, range, etc, as well as the human playing it. More complexity to add to such a software.

• Talking about standards falling short. Imagine a guitar: you can pluck the strings near the bridge or near the sound hole ("Sul Ponticello" & "Sul Tasto"), obtaining different sounds. if you have a hexaphonic midi pickup and the "right" guitar to midi converter it is theoretically possible to detect automatically if you are playing near the bridge or the hole (the genious that developed the algorythm got United States Patent 5717155 ). But....midi doesn't support natively that information (one of the many possibles articulations, together with harmonics, staccatto, etc,etc). So using midi as an intermediate step will result in that information being lost (or you will need to develop a custom, ugly workaround for transporting the articulation info in midi, and then reading into a software that articulation ( the software must be configured for your solution of transporting articulation info).

Because of all these reasons, I believe it will take a while before we can convert an mp3 and get a decent score (unless it is a very simple tune).

My first dream is first to see a notation standard (like musicxml) supporting ALL notation: from ALL the classical, romantic and post romatic period scores, to ethnic middle east music: meterless passages, microtonal music, all cases of "weird" notation: like feathered beams, cross staff beams, osias and advanced footnotes that combine text and notation, advanced notation with grace notes, all forms of slurs, let ring-, extended slurs, splayed stems, alternate beaming, different ways of notating harmonics, all guitar notation for bending, etc,etc,etc).

My second dream is to see a standard for articulations in midi (or other protocol). There have been attempts to standardize articulation ( e.g. Spitfire's UACC -- Universal midi CC Articulation CC system, or Expressions maps,). But so far nothins has become a standard.

Ariel//