Piano World Home Page Forums Our Most Popular Forums Digital Pianos - Electronic Pianos - Synths & Keyboards Any advances-- converting mp3 or wav file to midi files?

i think this subject was approached a couple years ago and the consensus was there wasn't anything good for converting mp3 or wave data into a midi file - that could be read as notation.

Is that still the case, or has anything better come out since then?

one of my favorite solo piano works is Dave Grusin's Random Hearts- on his solo piano movie theme CD. No sheet music or lead sheets available. a little too complex to figure out in my head and i'm not a transcriber. so the mp3 to midi - even if crappy- would give me a foundation to work off.

anything interesting out there to play with?

thanks!

widisoft WIDI, maybe? I haven't used it, only heard.

that seems to be first thing pops up when googled. wonder if its really worth 129 bucks!

Maybe try a free software implementation first. It might be good enough since there is only one timbre to worry about hopefully.

Or maybe someone else in the forum has mor knowledge about this.

I've tried Audio Score & Intelliscore, not very accurate.
Hit and miss .... you have to spend so much time cleaning up the notation. it's not worth the time.

IMO ... use something like Transcribe from www.seventhstring.com, really good for figuring out songs.

Probably many softwares can do the the same thing, but this one allows you to set up your own shortscuts for making selections, loops, etc ....

Works great with a foot pedal also.

what i was hoping to accomplish with it it to at least see enough in the output to be able to discern a melody line and identify a chord structure underneath. and then take that, create a lead sheet from it and then reconstruct from there. i think i'll try something free and see if i can get anywhere close....

dannac- thanks for that website- really looks interesting. they pretty much say the "problem" has not yet been solved...

Bruce, you may also wish to try the trial version of Melodyne in order to decipher the notes of the tune:

http://www.youtube.com/watch?v=I4YEebBN2ok

Cheers,
James
x

thanks James.... Now that really is fascinating stuff. appreciate your pointing it out to me. a bit outside my price range for what i want to do, but really technologically interesting.

i guess the beauty of technology like that is it might allow someone with more musical talent in their head than in their hands to be able to fully express themselves.

Yes, I believe there is some truth in that.

However, the cynic in me suggests that this kind of software may actually result in more over-produced, sterile music, where everything is beat-for-beat and note-for-note perfect.

James
x

its not cynicism- its the truth. music quantization and midi editing is like that Christian Bale movie - Equilibrium. no highs, no lows... no life!

this one program- amazing midi- might be worth looking at. and if that doesn't work. i guess i will just have to listen to the song long enough to where i can work it out in my head because i can see this spinning off course into a techno-trap quickly.

My take on this is that if each key/note of a piano only produced the fundamental (no overtones at all), it could be possible. But that is not the case.

Even the simplest of cases (hypothetical of course) where each note had only the fundamental and the first overtone (an octave higher), and the pianist struck only two keys an octave apart (dampers on the strings to prevent SR), how could software tell if one or two or three keys were struck?

Now add the complexity of all the possible overtones in one string (twelve or so will pretty well define the timbre), and no dampers on the strings.

Now suppose that the piano is playing middle C, and so is the clarinet. The fundamentals of both are pure sine waves which have identical shape (it's only the partials that define timbre). I'm not betting that software can tell them apart.

I'm not a believer that this one will be sorted out soon by software.

What then becomes so amazing is that gifted and experienced musicians can do it.

Glenn

If you like digging into papers and do research, I would recommend you looking into this:

http://www.music-ir.org/mirex/wiki/

To be honest - with any complication with the input recording, it is not solvable and I think won't be.
Bach Inventions - possibly; Chopin's Fantasie-Impromptu - forget about it.

seems like the same subject i want to talk about !! is Is it possible to convert CD , mp3 song files to midi and using a program software like - PreSonus Studio One, when it has been converted to midi take out the sounds you want, e.g In my case I would leave the drums and bass and that would be like my backing track ?

In theory, yes.
In practise, it's not quite as successful as you would hope.

The quality of the outcome depends largely on the input audio source. If you are feeding the audio->MIDI software a fast, complicated track, with lots of instruments, it's unlikely that you will get a 'clean', accurate MIDI representation.

James
x

And reverb, whether naturally occurring or added in post-processing, seems to make it hugely more difficult.

Hi,

This problem has a parallel in the visual realm called "image recognition" software. Let's say you wanted the computer to make a positive ID of all the trees it sees with it's on-board camera. I believe most cutting-edge research software can make positive IDs of trees from photos or a still from a well-shot video. But if you increase the noise with FX filters or show it a loose impressionist painting of a tree, the computer at this stage of the game, in 2012, will fail. Software is not able to always filter the unnecessary information and get down to the core of what 'a tree' is and the general Platonic concept of a tree that us humans have in our minds.

There are 2 ways to deal with this problem - brute force and machine learning. Brute force would be to give it a huge database of every image of a tree you can find(close to an infinite variety). Machine learning would be to teach it like teaching a small child - 'this is tree, this is what a tree typically contains (leaves, a trunk, etc.)' and then through agonizing trial and error, revision after revision, you might get what we have today in cutting-edge software.

A practical example of fooling today's software is when you have to enter a secret code on a website from an image that has been distorted somehow. It might read "Apple" but the letters are skewed and lower software cannot handle this yet. But humans right away say, "oh that says Apple. "

Sorry to diverge there, but getting back to "acoustic recognition" software, the same brute force or machine learning techniques are applicable. If you wanted the computer to recognize a C minor chord no matter what inversion or what instrument(s) are playing it, you could give it a huge database of every possible C minor chord instance ever written in classical/jazz literature and then it would just compare the chord in question to its HUGE database of harmonies. Or, and this is the way of the future I think, you start from scratch and teach it basic ear-training. To my knowledge, this is not yet possible in 2012 - to accurately notate harmonies and separate the individual notes that make them up, and then transfer that into a score, one after the other.

I think software could be taught (and maybe has already been taught) how to hear melody in real time, as it is linear. But harmony and its subtleties is much more difficult (even for humans with Absolute Pitch). It will come someday, but it might be another 10 years from now (when we have more massive parallel processing and 3D chip technology).

That being said, we are living in a great time for technology advances just on the horizon!

What you are saying touches the core of the problem. Image recognition programs can address very specific questions like "detect if a tree is on a picture" or "find a licence plate on an image and if there is, what is written on it", but not "tell me what's on the picture".

If you would like to apply that to music, a relevant question that can be answered would be "in this piece, does the clarinet play c or d note", and the one that is too broad would be "name all the instruments and write down what they are playing"

Yes I agree that a computer could be taught to pick out a pitch and match the Hz to its table of frequencies. And like you said, to be able to pick out an instrument in a dense orchestral chord and say what note it is playing - that's way off down the road yet technologically speaking.

I was pondering this afternoon how this problem might be solved and there may be a way through harmonic/spectral analysis. All clarinets for example have a certain dominance of harmonics, as does a violin or a trumpet. If we ('we' meaning computer scientists) could teach a computer to identify timbre, it would be able to make an educated guess as to what instrument it is listening to and what pitch it is playing (by subtracting the unnecessary overtones for that particular instrument's spectral signature).

As Glenn already mentioned above, when you get a wash of harmonics like a piano with the sustain pedal down, you would need serious algorithms to try an subtract the unnecessary overtones and then try to single out the individual pitches. I think this will ultimately require a hybrid approach of brute force (a huge table of a piano's harmonic spectra to refer to) as well as some decent machine learning (ear training for quality of Major vs. minor chords, inversions, octave doublings, etc.)

I'm not saying it is impossible - everything that can be computed in this universe will ultimately be computed by the time we humans are done with our quest for knowledge/tech. But I wouldn't know how to even start this kind of software project. It maybe a decade before we make significant strides.

Which stinks because I would love to have Bill Evans' complete solo piano recordings transcribed note for note in MIDI, and with his dense lush harmonies, it would take me forever to do just one album by ear.

I think the majority of Bill Evans arrangements (or at least a good many of them) have been transcribed onto sheet music and commercially published.