Show HN: I trained a 125M model to autocomplete piano on-device

194

ssimedw about 3 hours ago 45 commentsRead Article on simedw.com

HI version is available. Content is displayed in original English for accuracy.

I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15).

The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device.

The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.

⚡ Community Insights

Discussion Sentiment

91% Positive

Analyzed from 869 words in the discussion.

Discussion (45 Comments)Read Original on HackerNews

goda90•about 1 hour ago

Reminds me of this project to generate every melody possible algorithmically in order to fight music copyright lawsuits. https://allthemusic.info/

jasonjmcghee•16 minutes ago

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

Tepix•16 minutes ago

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).

timmb•about 2 hours ago

Reminds me of Francois Pachet’s Continuator (all the way back in 2003, using hierarchical markov models)

https://www.francoispachet.fr/continuator/

hliyan•about 1 hour ago

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.

killerstorm•40 minutes ago

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.

HPsquared•29 minutes ago

I'm thinking canyon.mid on Microsoft GS Wavetable Synth.

xtracto•25 minutes ago

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

heikkilevanto•about 2 hours ago

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.

Naitik88•21 minutes ago

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.

vunderba•41 minutes ago

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

morkalork•about 2 hours ago

That's a fun idea. You could start playing the piano and it kicks in with a base and drums for a jazz band.

gaya3bollineni•about 1 hour ago

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?

simedw•36 minutes ago

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.

Naitik88•23 minutes ago

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.

devonsolomon•about 2 hours ago

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.

Naitik88•23 minutes ago

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.

mh-•about 1 hour ago

+1 all of this. That would be incredible (this already seems very cool - excited to get home and try it!)

davidajackson•about 1 hour ago

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...)

speedgoose•about 1 hour ago

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

evalystai•about 3 hours ago

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).

simedw•about 2 hours ago

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.

subhajeet2107•about 2 hours ago

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

simedw•about 1 hour ago

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.

isoprophlex•about 2 hours ago

Amazing idea! Gonna hook this up to my little synthesizer and blast some square wave arpeggiated ML music!

leobg•about 2 hours ago

I don’t have MIDI. How about whistling or playing the piano via microphone? Sounds easy. Another 6 month rabbit hole? :)

pil0u•about 2 hours ago

Even after a few years deep into AI, I find your application absolutely magic. This is very inspiring, thank you for sharing.

the_black_hand•42 minutes ago

Cool work. I tried using LLMs to parse sheet music and they are really bad.

aghilmort•26 minutes ago

really incredible work! great use case, impeccable learning strategy, congrats!

arkmm•about 1 hour ago

Very cool! Can you say a little bit about the size of the DPO training examples and how long training took?

simedw•about 1 hour ago

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.

mandeepj•22 minutes ago

Gemma 4 E2B was too heavy for your needs?

yoouareperfect•about 3 hours ago

This is really awesome thanks for sharing

jwr•about 1 hour ago

> Eventually I used Gemini 3.5 Flash for pairwise evaluation

But, but… wouldn't that be… (gasp) DISTILLATION?

Fun project!