I think this is a great project and very HN. Not sure why the comments are so focused on the deliverable- you learned way more and had a much more interesting experience.
One think I didn't see mentioned in the post- maybe I missed it- how large was the data? How many samples did you use to pretrain and post-train
You can also listen to the transcript of four Russian composers, including Rachmaninoff, playing this pattern recognition and generation game at a dinner party in the late 1800’s: https://youtu.be/PlFPOWuwBHI?is=EKBK7QQkJs4MsTCU
Composers at the time could do this just by looking at sheet music and audiating, without using a piano.
To add to this: classical extemporization, post-Romantic period, also sort of fell out of favor as the composer began to pre-empt the performer and the "written note is sacrosanct" effectively became the de jure standard.
There's a great story around how Beethoven, perhaps one of the strongest improvisers of his day, completely upstaged Steibelt, a contemporary musician and by all accounts a bit of a charlatan. I'll include the entire quote verbatim from "The Lives of the Great Pianists" by Harold Schonberg.
This time Steibelt had prepared a brilliant fantasy for piano and strings, the theme coming from the Beethoven trio he had heard the previous week. Steibelt's admirers were in raptures. Now the issue was joined, and Beethoven had to show his strength. He walked to the piano, grasped the cello part of the Steibelt work en route, put it upside down on the piano and insultingly drummed out a theme with one finger. Then he improvised, and—angry, excited, on his mettle—how he must have improvised! Before Beethoven had finished, Steibelt stole from the room. He never again would meet Beethoven and made it a condition before going anywhere in Vienna that Beethoven not be invited.
I would love something like that, except that I play the melody, and it produces proper 3-4 part accompaniment, preferably in good baroque style. Extra bonus if it could also write it into a file in a format suitable for music editing programs.
An early attempt at this was Microsoft Songsmith [1] all the way back in 2009, which would take a melody (usually recorded by mic) and try to scaffold an accompaniment around it though obviously not realtime in any sense of the word.
The closest we've had to realtime orchestration around a melody in the "real world" is probably arranger keyboards though your left hand is still responsible for the chord progression itself.
This feels like a natural next step. Starting with a simple melody and having the system fill in the rest while still following your playing style could make it much more useful for experimentation than generating a complete piece from scratch.
That is an incredibly hard challenge though. Creating the whole backing track (in any meaningful way other than just basic chords) from just melody will require an amazingly high number of highly subjective choices and random gen will not lead to good outcomes since our ears like intentional and artistical creativity in general.
Interesting. I do feel letting a machine generating notes is taking the joy out of improvisation, is it not?
I think one of the great, early, joys of learning a piano is gaining the following intuitions: The seemingly harder path of learning sheet is actually faster. Your mind _should_ learn to think in two dimensions Spatial — where fingers go — and Time — pitch and tempo – ** when learning. The _internalization_ of Space and Time queues guide the fingers in a dance that is vastly satisfying. This skill leads you to the final part of the journey that is improvisation and the one more exciting than what i am on now.
---
**
Space: Your finger placement on keys right, e.g. knowing how to go from landmark/anchor notes(mid-c, G, F etc) and then go to the others above and below it. Crudely this is some what like typing from your landmark f and j qwerty keyboard
Time: The out singing/verbalizing of the notes/beats on a time measure as you play them(per the time measure). e.g. you can say out loud 1-2-3-4 for 4/4 measure, if the measure has quarter notes say out loud. And `1-e-and-a-2-e-and-a-3-e-and-a-4-e-and-a` for a 4/4 with 1/16th note granularity. Do this as you play the notes and you get a sense of tempo.
Ah, MIDI files. The only type of music you could realistically download from the internet back in the day, and you had to wake up at ungodly hours so that your dialup modem would not rack up a massive phone bill.
I remember at one point RuneScape switched from the built in Microsoft midi whatever to their own sound engine, and from that day, everything sounded wrong, even the frogs, because to me, the crappy midi sounds were the whole personality and feel of the game.
This is really fun. Scaler 3 starts with a chord progression and lets you break it down into musical performances and parts. Useful for ideation when producing.
Would be fun to get a midi clock going and play some chords on my piano and have my synth start jamming along with the bass and my keyboard doing some performance. Or any combination of the above.
That would be a really interesting direction. At that point it starts feeling less like autocomplete and more like having another musician reacting to what you're playing in real time.
The idea is awesome! :) However there's definitely much room for improvement, first of all rythm and composition (so there's some sense of musical form).
Yes, I think I’ve gotten it to roughly a GPT-2 level: good enough to share, but with a lot of room left to improve. I think adding some kind of bar/measure token might help with rhythm, and perhaps some form of longer-term planning for the overall composition.
How would you expand this to support elements like attack ("velocity of the key-down" in piano speak), grace notes, timing etc. Would each of those be part of this model or another model? How would you model an arbitrary element (pedal, duration, etc...)
The biggest speed improvement came from changing the note representation when I switched to compound note events: roughly 5× fewer autoregressive passes per note.
For the current model I’m using Core ML, which optimizes the kernels the first time you run it. I haven’t actually spent that much time tuning performance beyond that.
The answer about changing the note representation was interesting. Sometimes a change in how the problem is represented ends up giving a much bigger improvement than trying to optimize the model itself.
Oh, a cool idea! I just tried it, works pretty well. Kudos!
One feature request:
Instead of playing the AI-generated audio solely through the iPhone's speakers, add an option to send the audio as midi notes to a device (probably the same one you received the mini notes from).
This is so amazing, can you improve the quality of generation at the cost of notes per seconds ? No one can play 108 notes/sec anyways, maybe you can train the model to do CoT for better quality
Yes, some kind of planning step is on my TODO list. Another thing I want to try is generating a few continuations in parallel, picking the one that looks best, and then continuing from there. Maybe the picking could be automatic.
I can probably squeeze out quite a bit more than 100 notes/sec as well. I haven’t spent much time optimizing inference yet.
Luckily I have access to 4x RTX 4090s, so I didn’t have to pay cloud GPU prices directly. If I had, it probably would have added up quite a bit given how many training runs and experiments I ended up doing.
This is honestly astonishing and the first "AI music" I've heard that has the potential to sound beautiful. I always thought that MIDI would be a perfect format for this. Glad to see this person make it happen!
I think this is a great project and very HN. Not sure why the comments are so focused on the deliverable- you learned way more and had a much more interesting experience.
One think I didn't see mentioned in the post- maybe I missed it- how large was the data? How many samples did you use to pretrain and post-train
This sort of “autocomplete” is actually fundamental to how classical composers were trained.
For anyone interested, I’d highly recommend reading Robert Gjerdingen’s article Gebrauchs-Formulas. https://www.researchgate.net/publication/259731561_Gebrauchs...
You can also listen to the transcript of four Russian composers, including Rachmaninoff, playing this pattern recognition and generation game at a dinner party in the late 1800’s: https://youtu.be/PlFPOWuwBHI?is=EKBK7QQkJs4MsTCU
Composers at the time could do this just by looking at sheet music and audiating, without using a piano.
To add to this: classical extemporization, post-Romantic period, also sort of fell out of favor as the composer began to pre-empt the performer and the "written note is sacrosanct" effectively became the de jure standard.
There's a great story around how Beethoven, perhaps one of the strongest improvisers of his day, completely upstaged Steibelt, a contemporary musician and by all accounts a bit of a charlatan. I'll include the entire quote verbatim from "The Lives of the Great Pianists" by Harold Schonberg.
Hearing the start of Für Elise, and then it being taken in an incredibly different direction, is surprisingly disconcerting.
Can we hear it?
It's on the linked page if you scroll down just a bit.
Reminds me of this project to generate every melody possible algorithmically in order to fight music copyright lawsuits. https://allthemusic.info/
Reminds me of Francois Pachet’s Continuator (all the way back in 2003, using hierarchical markov models)
https://www.francoispachet.fr/continuator/
I would love something like that, except that I play the melody, and it produces proper 3-4 part accompaniment, preferably in good baroque style. Extra bonus if it could also write it into a file in a format suitable for music editing programs.
An early attempt at this was Microsoft Songsmith [1] all the way back in 2009, which would take a melody (usually recorded by mic) and try to scaffold an accompaniment around it though obviously not realtime in any sense of the word.
The closest we've had to realtime orchestration around a melody in the "real world" is probably arranger keyboards though your left hand is still responsible for the chord progression itself.
[1] - https://en.wikipedia.org/wiki/Microsoft_Research_Songsmith
This feels like a natural next step. Starting with a simple melody and having the system fill in the rest while still following your playing style could make it much more useful for experimentation than generating a complete piece from scratch.
That is an incredibly hard challenge though. Creating the whole backing track (in any meaningful way other than just basic chords) from just melody will require an amazingly high number of highly subjective choices and random gen will not lead to good outcomes since our ears like intentional and artistical creativity in general.
That's a fun idea. You could start playing the piano and it kicks in with a base and drums for a jazz band.
Question as a musician: Do you find satisfaction in the piece it created in-and-of-itself (subtracting the satisfaction from making the tool itself)?
Interesting. I do feel letting a machine generating notes is taking the joy out of improvisation, is it not?
I think one of the great, early, joys of learning a piano is gaining the following intuitions: The seemingly harder path of learning sheet is actually faster. Your mind _should_ learn to think in two dimensions Spatial — where fingers go — and Time — pitch and tempo – ** when learning. The _internalization_ of Space and Time queues guide the fingers in a dance that is vastly satisfying. This skill leads you to the final part of the journey that is improvisation and the one more exciting than what i am on now.
---
**
Space: Your finger placement on keys right, e.g. knowing how to go from landmark/anchor notes(mid-c, G, F etc) and then go to the others above and below it. Crudely this is some what like typing from your landmark f and j qwerty keyboard
Time: The out singing/verbalizing of the notes/beats on a time measure as you play them(per the time measure). e.g. you can say out loud 1-2-3-4 for 4/4 measure, if the measure has quarter notes say out loud. And `1-e-and-a-2-e-and-a-3-e-and-a-4-e-and-a` for a 4/4 with 1/16th note granularity. Do this as you play the notes and you get a sense of tempo.
OP found a different source of joy: "I wanted the fun of working through the problem myself, rather than just implementing someone else’s research."
Fair enough.
Ah, MIDI files. The only type of music you could realistically download from the internet back in the day, and you had to wake up at ungodly hours so that your dialup modem would not rack up a massive phone bill.
MIDI is still widely used for professional music production. It sounded goofy back in the day because synthesizer it was played on was not very good.
I remember at one point RuneScape switched from the built in Microsoft midi whatever to their own sound engine, and from that day, everything sounded wrong, even the frogs, because to me, the crappy midi sounds were the whole personality and feel of the game.
I'm thinking canyon.mid on Microsoft GS Wavetable Synth.
I remember downloading MOD files. It was just a bit larger than midi, but sounded better. I had a PC but I think it was an Amiga thing
Need For Madness is one of my favorite games, and probably two thirds of that fondness is the music, which was a bunch of MOD files.
Congratulations! It is amazing!!! Can you do the same with a song and give different drums to see which one fits better?
Für Elsevier Journal Access Library pass out candy for safety this halloween your baby off milk shake it off.
I, für one, welcome our new LLM overlords.
This is really fun. Scaler 3 starts with a chord progression and lets you break it down into musical performances and parts. Useful for ideation when producing.
Would be fun to get a midi clock going and play some chords on my piano and have my synth start jamming along with the bass and my keyboard doing some performance. Or any combination of the above.
That would be a really interesting direction. At that point it starts feeling less like autocomplete and more like having another musician reacting to what you're playing in real time.
+1 all of this. That would be incredible (this already seems very cool - excited to get home and try it!)
The idea is awesome! :) However there's definitely much room for improvement, first of all rythm and composition (so there's some sense of musical form).
Thank you.
Yes, I think I’ve gotten it to roughly a GPT-2 level: good enough to share, but with a lot of room left to improve. I think adding some kind of bar/measure token might help with rhythm, and perhaps some form of longer-term planning for the overall composition.
How would you expand this to support elements like attack ("velocity of the key-down" in piano speak), grace notes, timing etc. Would each of those be part of this model or another model? How would you model an arbitrary element (pedal, duration, etc...)
Running a 125M model on-device at that speed is impressive. How much did you have to optimize the model to get that performance on an iPhone?
The biggest speed improvement came from changing the note representation when I switched to compound note events: roughly 5× fewer autoregressive passes per note.
For the current model I’m using Core ML, which optimizes the kernels the first time you run it. I haven’t actually spent that much time tuning performance beyond that.
The answer about changing the note representation was interesting. Sometimes a change in how the problem is represented ends up giving a much bigger improvement than trying to optimize the model itself.
Amazing idea! Gonna hook this up to my little synthesizer and blast some square wave arpeggiated ML music!
Oh, a cool idea! I just tried it, works pretty well. Kudos!
One feature request:
Instead of playing the AI-generated audio solely through the iPhone's speakers, add an option to send the audio as midi notes to a device (probably the same one you received the mini notes from).
This is so amazing, can you improve the quality of generation at the cost of notes per seconds ? No one can play 108 notes/sec anyways, maybe you can train the model to do CoT for better quality
Yes, some kind of planning step is on my TODO list. Another thing I want to try is generating a few continuations in parallel, picking the one that looks best, and then continuing from there. Maybe the picking could be automatic.
I can probably squeeze out quite a bit more than 100 notes/sec as well. I haven’t spent much time optimizing inference yet.
Talking about AI music with some live human MIDI inputs, Magenta Realtime 2 was released a few weeks ago and is pretty fun.
https://magenta.withgoogle.com/magenta-realtime-2
Even after a few years deep into AI, I find your application absolutely magic. This is very inspiring, thank you for sharing.
I don’t have MIDI. How about whistling or playing the piano via microphone? Sounds easy. Another 6 month rabbit hole? :)
Should tokens be multiple Midi notes? Why or why not?
Are the weights of the model available?
I too am currently learning to play Sonatina in G Major, it's a fun one! the end is surprisingly tricky
Would you be willing to share about how much it cost to train a model like this?
Luckily I have access to 4x RTX 4090s, so I didn’t have to pay cloud GPU prices directly. If I had, it probably would have added up quite a bit given how many training runs and experiments I ended up doing.
This is honestly astonishing and the first "AI music" I've heard that has the potential to sound beautiful. I always thought that MIDI would be a perfect format for this. Glad to see this person make it happen!
Very cool! Can you say a little bit about the size of the DPO training examples and how long training took?
For DPO I only had around 700 preference examples, so not much data at all. That took about 12 minutes to train on a single GPU.
Pretraining was obviously a a lot slower, the 125M model took roughly half a day.
Pretty Impressive!!
Cool work. I tried using LLMs to parse sheet music and they are really bad.
This is really awesome thanks for sharing
really incredible work! great use case, impeccable learning strategy, congrats!
> Eventually I used Gemini 3.5 Flash for pairwise evaluation
But, but… wouldn't that be… (gasp) DISTILLATION?
Fun project!
The horror!
Gemma 4 E2B was too heavy for your needs?