> I trained a 125M model to autocomplete piano on-device
I first parsed the "on device" as _on the piano_ and was intrigued how that worked :) maybe someone can make "dynamic" rolls for a player piano somehow (like one of those braille "screens")?
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.
Classical pianist and software product designer here.
I see so much in common with this project and the numerous AI-based UX design tools out there. Whether it's music or UI, now that the "generation" portion of the work costs zero, all that remains is taste.
And so much of taste comes from exploring and killing off possibilities that turn out to be dead-ends. I love the idea that models like these will help us find the dead ends faster, or even produce a gem here and there.
P.S. if you want another uncanny version of Fur Elise, listen to Beethoven's own 1822 revision: https://www.youtube.com/watch?v=s24TtiGgb6k. His 1810 version that we all know was simpler and more balanced. But for what it's worth, Beethoven didn't publish either of them.
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
The disconcerting effect is enhanced by hearing the Dutch voice and reading the Dutch subtitles which don't quite match up with what he is saying in English! - until it all goes off into musical absurdity. Thanks for sharing.
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.
But most people who want to get good at their main discipline, even for self-fulfillment purposes, have been gaining the necessary experience through their job, because an unrelated job takes away too much time to do so.
its a good motivator and looks like we'll be returning to the more feudal version where just that class was focusing on the arts, only because everyone else is preoccupied with doing the things that pays the bills
unless they're actually serfs again and won't have bills to pay, but no access to anything outside of the fiefdom that they maintain all day
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.
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.
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 results strike me as comparable or worse than you could get with a Markov model. I think it reveals the gap in understanding between LLMs and music. I think you need to either:
- Set up a pipeline to decompose music into, say, harmonic sequences and melodic sequences, and then have the LLM work on some more fundamental or more high-level layer of musical composition and then re-translate it back into actual sounds.
- Develop a better dataset and train the LLM more natively on musical examples.
Does anybody know of a project that has produced more convincing results?
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.
Looks neat! I'd love for it to feed the MIDI notes back into my player piano instead of playing out of my iPhone's (comparatively tinny) speakers though.
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).
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...)
It occurres to me that the relationship between pitches usually matters more than the pitches themselves. Perhaps the pitches could be encoded in a vector like what is typically done for position.
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.
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.
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 disgusting. I'm sorry. All the positive comments. It's like Waymo's posts on HN. Incomprehensible to me.
I guess no one actually wants to learn about harmony, about voicing, and voice leading, spend the hours. No one wants to learn how to actually play an instrument. I guess no one is willing to do the work.
They want to just press some keys and declare that they made what the computer generated.
This is all so depressing for me.
I think in the near future, the world will divide, and all of the people who wish to think and do, will put up a wall separating themselves from the hu-bots that infest the rest of the planet. It's alright I guess. I hope there isn't a slaughter of one side or the other.
>I guess no one actually wants to learn about harmony, about voicing, and voice leading, spend the hours. No one wants to learn how to actually play an instrument. I guess no one is willing to do the work. They want to just press some keys and declare that they made what the computer generated.
I don't understand that as a takeaway. Even if this worked perfectly, it would not be meaningfully stopping or discouraging anyone from learning music, and no one is claiming that what this models outputs is something the player played. It's a toy that someone made presumably because they like piano.
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 trained a 125M model to autocomplete piano on-device
I first parsed the "on device" as _on the piano_ and was intrigued how that worked :) maybe someone can make "dynamic" rolls for a player piano somehow (like one of those braille "screens")?
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.
I'm a little lost. What am I listening for in that video?
Classical jamming.
Classical pianist and software product designer here.
I see so much in common with this project and the numerous AI-based UX design tools out there. Whether it's music or UI, now that the "generation" portion of the work costs zero, all that remains is taste.
And so much of taste comes from exploring and killing off possibilities that turn out to be dead-ends. I love the idea that models like these will help us find the dead ends faster, or even produce a gem here and there.
P.S. if you want another uncanny version of Fur Elise, listen to Beethoven's own 1822 revision: https://www.youtube.com/watch?v=s24TtiGgb6k. His 1810 version that we all know was simpler and more balanced. But for what it's worth, Beethoven didn't publish either of them.
Thanks. I found the 1822 version actually quite agreable
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
There is this in the post:
> The final dataset contained a few hundred thousand MIDI files, representing roughly 300 million note events.
Hearing the start of Für Elise, and then it being taken in an incredibly different direction, is surprisingly disconcerting.
There's an incredibly funny bit (if you're into absurdist comedy) by Hans Teeuwen about the disconcerting direction a well known song can take:
https://www.youtube.com/watch?v=FCDPxNyLAww
The disconcerting effect is enhanced by hearing the Dutch voice and reading the Dutch subtitles which don't quite match up with what he is saying in English! - until it all goes off into musical absurdity. Thanks for sharing.
I found it refreshing and interesting, not really disconcerting. Show's how AI can extrapolate in novel ways. An interesting feature.
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/
Thanks. I enjoyed watching the TEDx talk on that page.
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.
Reminds me of a story a coworker told me about his band rehearsal when the new drummer counted one-two-three-four-five-six-se-ven-eight.
as with every discipline that AI and therefore computers can reason through suddenly:
you could always do the discipline you liked before for self fulfillment purposes, and you can still do it for self fulfillment purposes now
But most people who want to get good at their main discipline, even for self-fulfillment purposes, have been gaining the necessary experience through their job, because an unrelated job takes away too much time to do so.
its a good motivator and looks like we'll be returning to the more feudal version where just that class was focusing on the arts, only because everyone else is preoccupied with doing the things that pays the bills
unless they're actually serfs again and won't have bills to pay, but no access to anything outside of the fiefdom that they maintain all day
Really cool! It reminds me of a demo I saw live at NeurIPS: https://www.youtube.com/watch?v=8s3V922h3CU (paper at https://openreview.net/pdf?id=3yeBer3J5z).
If you like such creative AI work, I would recommend looking at the other "NeurIPS Creative AI Track 2025" submissions as well.
Reminds me of Francois Pachet’s Continuator (all the way back in 2003, using hierarchical markov models)
https://www.francoispachet.fr/continuator/
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.
MIDI is a protocol, and it's not ever going away or being replaced.
Some protocols (like I2C or MIDI) are gonna be with humanity forever, probably.
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.
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's a fun idea. You could start playing the piano and it kicks in with a base and drums for a jazz band.
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.
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.
+1 all of this. That would be incredible (this already seems very cool - excited to get home and try it!)
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.
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.
The results strike me as comparable or worse than you could get with a Markov model. I think it reveals the gap in understanding between LLMs and music. I think you need to either: - Set up a pipeline to decompose music into, say, harmonic sequences and melodic sequences, and then have the LLM work on some more fundamental or more high-level layer of musical composition and then re-translate it back into actual sounds. - Develop a better dataset and train the LLM more natively on musical examples.
Does anybody know of a project that has produced more convincing results?
There is no LLM in the ML pipeline provided by the OP.
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.
Looks neat! I'd love for it to feed the MIDI notes back into my player piano instead of playing out of my iPhone's (comparatively tinny) speakers though.
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).
You could also distill the fabulous Anticipatory Music Transformer from Stanford. https://crfm.stanford.edu/2023/06/16/anticipatory-music-tran...
Also consider checking their decisions about representation, etc
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...)
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)?
Amazing idea! Gonna hook this up to my little synthesizer and blast some square wave arpeggiated ML music!
It occurres to me that the relationship between pitches usually matters more than the pitches themselves. Perhaps the pitches could be encoded in a vector like what is typically done for position.
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.
I would love to see a jamming partner. So I could play along him on the same piece.
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.
https://www.youtube.com/watch?v=wyo3JDMsWyM
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? :)
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
This is almost exactly what Jordan Rudess is doing now with some folks from MIT or Stanford right now. Google his interview with Rick Beato.
Congratulations! It is amazing!!! Can you do the same with a song and give different drums to see which one fits better?
I wonder if training SOTA models on this will make them more humane?
next step is a model that autocompletes the part where you actually practice. mine just sits there judging my scales.
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.
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 really awesome thanks for sharing
Cool work. I tried using LLMs to parse sheet music and they are really bad.
Should tokens be multiple Midi notes? Why or why not?
Are the weights of the model available?
This is disgusting. I'm sorry. All the positive comments. It's like Waymo's posts on HN. Incomprehensible to me.
I guess no one actually wants to learn about harmony, about voicing, and voice leading, spend the hours. No one wants to learn how to actually play an instrument. I guess no one is willing to do the work.
They want to just press some keys and declare that they made what the computer generated.
This is all so depressing for me.
I think in the near future, the world will divide, and all of the people who wish to think and do, will put up a wall separating themselves from the hu-bots that infest the rest of the planet. It's alright I guess. I hope there isn't a slaughter of one side or the other.
>I guess no one actually wants to learn about harmony, about voicing, and voice leading, spend the hours. No one wants to learn how to actually play an instrument. I guess no one is willing to do the work. They want to just press some keys and declare that they made what the computer generated.
I don't understand that as a takeaway. Even if this worked perfectly, it would not be meaningfully stopping or discouraging anyone from learning music, and no one is claiming that what this models outputs is something the player played. It's a toy that someone made presumably because they like piano.
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!
> Eventually I used Gemini 3.5 Flash for pairwise evaluation
But, but… wouldn't that be… (gasp) DISTILLATION?
Fun project!
The horror!
really incredible work! great use case, impeccable learning strategy, congrats!
Pretty Impressive!!
I too am currently learning to play Sonatina in G Major, it's a fun one! the end is surprisingly tricky
Gemma 4 E2B was too heavy for your needs?