Music Generation with Magenta: Using Machine Learning in Arts


This is the text transcript of the presentation Music Generation with Magenta: Using Machine Learning in Arts, made at Devoxx Belgium 2019.


Generative Music

“Generative art is an artwork partially or completely created by an autonomous system.” This definition is good, but not really fun. Let’s other ways of seeing generative music.

Make music without being a musician

Maybe you don’t know how to improvise, maybe you need help to compose. Either way, a generative algorithm can help you do it.

By interacting with a machine, you can find new ideas, discover new sounds, or do even things you cannot by yourself. The machine becomes an extension of our creativity.

“Why you should build silly things” - Monica Dinculescu

It is okay to make art without taking ourselves seriously. Generative music is a good way of doing that, because those kind of systems can be interacted with easily.

Monica Dinculescu - Why you should build silly things

Helping people build generative systems

To make generative art, you need autonomous systems that makes it possible. As software developers, this is our role to provide that, whether it is for an art exhibit, a generative radio, or a fun website.

“The weird and the strange is good” - Brian Eno

Generative systems makes tons of mistakes (also humans), but mistakes are good. They form the basic mechanism with which we learn and are important in the creation process.

Representation: MIDI

MIDI is a musical representation analogous to sheet music, where note has a pitch, velocity, and time.

Working with MIDI shows the underlying structure of the music, but doesn’t define the actual sound, you’ll need to use instruments (numeric or analogic).

MIDI diagram

Representation: audio

Working with audio is harder because you have to handle 16000 samples per seconds (at least) and keep track of the general structure. Generating audio is more direct than MIDI.

It is more direct because you don’t need the extra step of using instruments to play the score.

Audio diagram

Machine Learning

Hand crafting the rules of a painting or the rules of a music style might be a hard task. That’s why Machine Learning is so interesting in arts: it can learn complex functions.

Music generation with RNNs (MIDI)

Inline - RNN diagram

Recurrent Neural Networks (RNNs) solves two important properties for music generation: they operate on sequences for the inputs and outputs and they can remember past events.

Long-term structure with LSTMs (MIDI)

Inline - LSTM diagram

Most RNN uses Long Short-Term Memory (LSTM) cells, since by themselves, RNNs are hard to train because of the problems of vanishing and exploding gradient, making long-term dependencies hard to learn.

By using input, output and forget gates in the cell, LSTMs can learn mechanisms to keep or forget information as they go.

Latent space interpolation with VAEs (MIDI)

Inline VAE diagram

Variational Autoencoders (VAEs) are a pair of networks where an encoder reduces the input to a lower dimensionality (latent space), from which a decoder tries to reproduce the input.

The latent space is continuous and follows a probability distribution, meaning it is possible to sample from it. VAEs are inherently generative models: they can sample and interpolate (smoothly move in the latent space) between two points.

Audio generation with WaveNet Autoencoders (audio)

Wavenet diagram

WaveNet is a convolutional neural network (CNN) taking raw signal as an input and synthesizing output audio sample by sample.

The WaveNet Autoencoder present in Magenta is a Wavenet-style AE network capable of learning its own temporal embedding, resulting in a latent space from which is it possible to sample and mix elements.

What’s in Magenta?

Here are some networks we’ll be using in this presentation and live code. Not all Magenta models are listed, but the main ones for music generation are.

Model Network Repr. Encoding
DrumsRNN LSTM MIDI polyphonic-ish
MelodyRNN LSTM MIDI monophonic
PolyphonyRNN LSTM MIDI polyphonic
PerformanceRNN LSTM MIDI polyphonic, groove
MusicVAE VAE MIDI multiple
NSynth Wavenet AE Audio -
GANSynth GAN Audio -

Live code: Generate a track

STEP 1: make everything sound like a cat.

We’ll use NSynth to mix cat sounds with other sounds.

STEP 1: The sounds

Audio diagram
Audio diagram
Audio diagram
Audio diagram

STEP 1: Entry point

See the file “” in the conference code and method app.

def app(unused_argv):
    encoding1, encoding2 = encode([FLAGS.wav1, FLAGS.wav2],
    encoding_mix = mix(encoding1, encoding2)
    synthesize(encoding_mix, checkpoint=FLAGS.checkpoint)

STEP 1: Encode

See the file “” in the conference code and method encode.

import numpy as np
from typing import List
from magenta.models.nsynth import utils
from magenta.models.nsynth.wavenet import fastgen

def encode(paths: List[str],
           sample_length: int = 16000,
           sample_rate: int = 16000,
          checkpoint: str = "checkpoints/wavenet-ckpt/model.ckpt-200000") \
  -> np.ndarray:
    audios = []
    for path in paths:
        audio = utils.load_audio(path, sample_length, sample_rate)
    audios = np.array(audios)
    encodings = fastgen.encode(audios, checkpoint, sample_length)
    return encodings

STEP 1: Mix

See the file “” in the conference code and method mix.

import numpy as np

def mix(encoding1: np.ndarray,
        encoding2: np.ndarray) \
    -> np.ndarray:
    encoding_mix = (encoding1 + encoding2) / 2.0
    return encoding_mix

STEP 1: Synthesize

See the file “” in the conference code and method synthesize.

import os
import time
import numpy as np
from magenta.models.nsynth.wavenet import fastgen

def synthesize(encoding_mix: np.ndarray,
               checkpoint: str = "checkpoints/wavenet-ckpt/model.ckpt-200000"):
    os.makedirs(os.path.join("output", "synth"), exist_ok=True)
    date_and_time = time.strftime("%Y-%m-%d_%H%M%S")
    output = os.path.join("output", "synth", f"{date_and_time}.wav")
    encoding_mix = np.array([encoding_mix])

STEP 1: GANSynth

As shown, the NSynth instrument is nice, but really slow for the audio synthesis. You should use GANSynth.

STEP 1: The results

Audio diagram Bass + Cat
Audio diagram Bass + Flute
Audio diagram Bass + Metal
Audio diagram Metal + Bass
Audio diagram Metal + Cat
Audio diagram Metal + Flute

STEP 2: sequence the cats

We’ll use DrumsRNN and MelodyRNN to generate MIDI to play the samples.

STEP 2: Reset

See the file “” in the conference code and method reset.

from magenta.protobuf.music_pb2 import NoteSequence

def reset(loop_start_time: float,
          loop_end_time: float,
          seconds_per_loop: float):
    sequence = NoteSequence()
    sequence = loop(sequence,
    return sequence

STEP 2: Loop

See the file “” in the conference code and method loop.

from import sequences_lib as ss
from magenta.protobuf.music_pb2 import NoteSequence

def loop(sequence: NoteSequence,
         loop_start_time: float,
         loop_end_time: float,
         seconds_per_loop: float):
    sequence = ss.trim_note_sequence(sequence,
    sequence = ss.shift_sequence_times(sequence,
    return sequence

STEP 2: Generate

See the file “” in the conference code and method generate.

from magenta.protobuf import generator_pb2
from magenta.protobuf.music_pb2 import NoteSequence

def generate(sequence: NoteSequence,
             name: str,
             bundle_filename: str,
             config_name: str,
             generation_start_time: float,
             generation_end_time: float):
    generator_options = generator_pb2.GeneratorOptions()
    generator_options.args['temperature'].float_value = 1
    sequence_generator = get_sequence_generator(name,
    sequence = sequence_generator.generate(sequence,
    sequence = ss.trim_note_sequence(sequence,
    return sequence

STEP 2: Sequence generator

See the file “” in the conference code and method get_sequence_generator.

import os
import as mm
from magenta.models.drums_rnn import drums_rnn_sequence_generator
from magenta.models.melody_rnn import melody_rnn_sequence_generator

def get_sequence_generator(name: str,
                           bundle_filename: str,
                           config_name: str):
    if name == "drums":
      generator = drums_rnn_sequence_generator
    elif name == "melody":
      generator = melody_rnn_sequence_generator
      raise Exception(f"Unknown sequence generator {name}")
    mm.notebook_utils.download_bundle(bundle_filename, "bundles")
    bundle = mm.sequence_generator_bundle.read_bundle_file(
        os.path.join("bundles", bundle_filename))
    generator_map = generator.get_generator_map()
    sequence_generator = generator_map[config_name](
        checkpoint=None, bundle=bundle)
    return sequence_generator

Wrapping up

This generative music demo helped us improvise and compose around an idea: a track composed of a percussion, a melody, and a cat 😺. We could interact with the system and it helped us improvise around a theme.

Can we do better?

Was it perfect? No, we had little happy accidents.

Inline - Bob Ross

Maybe we could have had more control over the sequences. Maybe we wanted to improvise around a musical style, or a specific structure.



The pre-trained models in Magenta are good, but if for example you want to generate a specific style, generate a specific time signature (3/4 for example), or a specific instrument (cello for example) you’ll need to train your own.

Datasets: LAKHS (MIDI)

A good place to start is the LAKHS dataset, a 180,000 MIDI files dataset, (partially) matched with the Million Song Dataset (metadata like artist, release, genre).

Datasets: NSynth (audio)

A large-scale and high-quality dataset of annotated musical notes. Training audio requires lots of resources, but can be achieved using GANSynth.

See the NSynth Dataset.

Building the dataset

From MIDI, ABCNotation, MusicXML files to NoteSequence.

convert_dir_to_note_sequences \
--input_dir="/path/to/dataset/jazz_midi/drums/v1" \
--output_file="/tmp/notesequences.tfrecord" \
Converted MIDI file /path/to/dataset/jazz_midi/drums/v1/TRVUCSW12903CF536D.mid.
Converted MIDI file /path/to/dataset/jazz_midi/drums/v1/TRWSJLM128F92DF651.mid.

Create SequenceExamples

The sequence examples are fed into the model, it contains a sequence of inputs and a sequence of labels that represents the drum track. Those are split to eval and training sets.

drums_rnn_create_dataset \
--config="drum_kit" \
--input="/tmp/notesequences.tfrecord" \
--output_dir="/tmp/drums_rnn/sequence_examples" \
[8,10): 1
[10,20): 8
[20,30): 8
[30,40): 29
[40,50): 1
[50,100): 2

Train and evaluate the model

Launch the training, using a specific configuration and hyperparameters.

drums_rnn_train \
--config="drum_kit" \
--run_dir="/tmp/drums_rnn/logdir/run1" \
--sequence_example_file="/tmp/drums_rnn/sequence_examples/training_drum_tracks.tfrecord" \
--hparams="batch_size=64,rnn_layer_sizes=[64,64]" \
Saving checkpoints for 0 into /tmp/drums_rnn/logdir/run1/train/model.ckpt.
Accuracy = 0.013341025, Global Step = 1, Loss = 6.2323294, Perplexity = 508.9396
Accuracy = 0.43837976, Global Step = 11, Loss = 5.1239195, Perplexity = 167.99252 (50.009 sec)
global_step/sec: 0.199963
Saving checkpoints for 12 into /tmp/drums_rnn/logdir/run1/train/model.ckpt.


You can launch TensorBoard and go to http://localhost:6006 to view the TensorBoard dashboard.

tensorboard --logdir="/tmp/drums_rnn/logdir"

Interaction with the outside world

Python to everything using MIDI

Magenta can send MIDI, which is understood by basically everything that makes sound: DAWs (like Ableton Live), software synthesizers (like fluidsynth), hardware synthesizers (though USB or MIDI cable), etc.

Magenta in the browser with Magenta.js

You can use Magenta and most of its models in the browser, using Magenta.js (which in turns uses Tensorflow.js).

Melody Mixer

RNN diagram

Melody Mixer

Neural Drum Machine

RNN diagram

Neural Drum Machine


RNN diagram


Easy peasy

        <!-- Load @magenta/music -->
        <script src="^1.0.0">
            // Instantiate model by loading desired config.
            const model = new mm.MusicVAE(
                '' +
            const player = new mm.Player();
            // Samples from MusicVAE and play the sample
            function play() {
                     .then((samples) => player.start(samples[0], 80));
    <body><button onclick="play()"><h1>Play Trio</h1></button></body>

Magenta in your DAW with Magenta Studio

Using Magenta.js and Max4Live (MaxMSP) process, Magenta Studio can be used directly in Ableton Live.


Dreambank - Can a machine dream?

Generative music using Magenta and Ableton Live, exhibit by Claire Malrieux.

Dreambank image

Hands-on Music Generation with Magenta

Upcoming book on Packt Publishing, expected publication date in January 2020.

Magenta book icon Packt logo