MusicLang : Controllable Symbolic Music Generation

🎶 You want to generate music that you can export to your favourite DAW in MIDI ?

🎛️ You want to control the chord progression of the generated music ?

🚀 You need to run it fast on your laptop without a gpu ?

Here is MusicLang Predict, your controllable music copilot.

I just want to try !

Go to our Colab, we have a lot of cool examples. From generating creative musical ideas to continuing a song with a specified chord progression.

I am more serious about it

Install the musiclang-predict package :

pip install musiclang_predict

Then open your favourite notebook and start generating music in a few lines :

from musiclang_predict import MusicLangPredictor
nb_tokens = 1024 
temperature = 0.9  # Don't go over 1.0, at your own risks !
top_p = 1.0 # <=1.0, Usually 1 best to get not too much repetitive music
seed = 16  # change here to change result, or set to 0 to unset seed

ml = MusicLangPredictor('musiclang/musiclang-v2') # Only available model for now

score = ml.predict(
    nb_tokens=nb_tokens,  # 1024 tokens ~ 25s of music (depending of the number of instruments generated)
    temperature=temperature,
    topp=top_p,
    rng_seed=seed # change here to change result, or set to 0 to unset seed
)
score.to_midi('test.mid') # Open that file in your favourite DAW, score editor or even in VLC

You were talking about controlling the chord progression ?

You had a specific harmony in mind am I right ? That's why we allow a fine control over the chord progression of the generated music. Just specify it as a string like below, choose a time signature and let the magic happen.

from musiclang_predict import MusicLangPredictor

# Control the chord progression
# Chord qualities available : M, m, 7, m7b5, sus2, sus4, m7, M7, dim, dim0.
# You can also specify the bass if it belongs to the chord (eg : Bm/D)
chord_progression = "Am CM Dm E7 Am" # 1 chord = 1 bar
time_signature = (4, 4) # 4/4 time signature, don't be too crazy here 
nb_tokens = 1024 
temperature = 0.8
top_p = 1.0
seed = 42

ml = MusicLangPredictor('musiclang/musiclang-v2')

score = ml.predict_chords(
    chord_progression,
    time_signature=time_signature,
    temperature=temperature,
    topp=top_p,
    rng_seed=seed # set to 0 to unset seed
)
score.to_midi('test.mid', tempo=120, time_signature=(4, 4))

Disclaimer : The chord progression is not guaranteed to be exactly the same as the one you specified. It's a generative model after all. Usually it will happen when you use an exotic chord progression and if you set a high temperature.

That's cool but I have my music to plug in ...

Don't worry, we got you covered. You can use your music as a template to generate new music. Let's continue some Bach music with a chord progression he could have used :

from musiclang_predict import MusicLangPredictor
from musiclang_predict import corpus

song_name = 'bach_847' # corpus.list_corpus() to get the list of available songs
chord_progression = "Cm C7/E Fm F#dim G7 Cm"
nb_tokens = 1024 
temperature = 0.8 
top_p = 1.0 
seed = 3666 

ml = MusicLangPredictor('musiclang/musiclang-v2')

score = ml.predict_chords(
    chord_progression,
    score=corpus.get_midi_path_from_corpus(song_name),
    time_signature=(4, 4),
    nb_tokens=1024,
    prompt_chord_range=(0,4),
    temperature=temperature,
    topp=top_p,
    rng_seed=seed # set to 0 to unset seed
)

score.to_midi('test.mid', tempo=110, time_signature=(4, 4))

What's coming next ?

We are working on a lot of cool features, some are already encoded in the model :

A control over the instruments used in each bar and their properties (note density, pitch range, average velocity)
Some performances improvements over the inference C script
A faster distilled model for real-time generation that can be embedded in plugins or mobile applications
An integration into a DAW as a plugin
Some specialized smaller models depending on our user's needs

How does that work ?

If you want to learn more about how we are moving toward symbolic music generation, go to our technical blog. The tokenization, the model are described in great details.

We are using a LLAMA2 architecture (many thanks to Andrej Karpathy awesome llama2.c), trained on a large dataset of midi files (The CC0 licensed LAKH). We heavily rely on preprocessing the midi files to get an enriched tokenization that describe chords & scale for each bar. The is also helpful for normalizing melodies relative to the current chord/scale.

Contributing & Contact us

We are looking for contributors to help us improve the model, the tokenization, the performances and the documentation. If you are interested in this project, open an issue, a pull request, or even contact us directly.

License

Specific licenses applies to our models. If you would like to use the model in your product, please contact us. We are looking forward to hearing from you ! MusicLang Predict is licensed under the GPL-3.0 License.

The MusicLang base language package on which the model rely (musiclang package) is licensed under the BSD 3-Clause License.

musiclang
/

musiclang-v2