MIDI-DDSP: Detailed Control of Musical Performance via Hierarchical Modeling

Demos | Blog Post
| Colab Notebook
| Paper | Hugging Face Spaces

MIDI-DDSP is a hierarchical audio generation model for synthesizing MIDI expanded
from DDSP.



You could install MIDI-DDSP via pip, which allows you to use the
cool Command-line MIDI synthesis to synthesize your MIDI.

To install MIDI-DDSP via pip, simply run:

pip install midi-ddsp


To train MIDI-DDSP, please first install midi-ddsp and clone the MIDI-DDSP repository:

git clone https://github.com/magenta/midi-ddsp.git

For dataset, please download the tfrecord files for the URMP dataset in
here to the data folder in
your cloned repository using the following commands:

cd midi-ddsp # enter the project directory
mkdir ./data # create a data folder
gsutil cp gs://magentadata/datasets/urmp/urmp_20210324/* ./data # download tfrecords to directory

Please check here for how to install and use gsutil.

Finally, you can run the script train_midi_ddsp.sh to train the exact same model we used in the paper:

sh ./train_midi_ddsp.sh

The current codebase does not support training with arbitrary dataset, but we will hopefully update that in the near

Side note:

If one download the dataset to a different location, please change the data_dir parameter in train_midi_ddsp.sh.

The training of MIDI-DDSP takes approximately 18 hours on a single RTX 8000. The training code for now does not support
multi-GPU training. We recommend using a GPU with more than 24G of memory when training Synthesis Generator in batch
size of 16. For a GPU with less memory, please consider using a smaller batch size and change the batch size
in train_midi_ddsp.sh.

Try to play with MIDI-DDSP yourself!

Please try out MIDI-DDSP
in Colab notebooks!

In this notebook, you will try to use MIDI-DDSP to synthesis a monophonic MIDI file, adjust note expressions, make pitch
bend by adjusting synthesis parameters, and synthesize quartet from Bach chorales.

We have trained MIDI-DDSP on the URMP dataset which support synthesizing 13 instruments: violin, viola, cello, double
bass, flute, oboe, clarinet, saxophone, bassoon, trumpet, horn, trombone, tuba. You could find how to download and use
our pre-trained model below:

Command-line MIDI synthesis

On can use the MIDI-DDSP as a command-line MIDI synthesizer just like FluidSynth.

To use command-line synthesis to synthesize a midi file, please first download the model weights by running:


To synthesize a midi file simply run the following command:

midi_ddsp_synthesize --midi_path <path-to-midi>

For a starter, you can try to synthesize the example midi file in this repository:

midi_ddsp_synthesize --midi_path ./midi_example/ode_to_joy.mid

The command line also enables synthesize a folder of midi files. For more advance use (synthesize a folder, using
FluidSynth for instruments not supported, etc.), please see synthesize_midi.py --help.

If you have a trouble downloading the model weights, please manually download
from here, and specify
the synthesis_generator_weight_path and expression_generator_weight_path by yourself when using the command line.
You can also specify your other model weights if you want to use your own trained model.

Python Usage

After installing midi-ddsp, you could import midi-ddsp in python and synthesize MIDI in your code.

Minimal Example

Here is a simple example to use MIDI-DDSP to synthesize a midi file:

from midi_ddsp import synthesize_midi, load_pretrained_model

midi_file = 'ode_to_joy.mid'
# Load pre-trained model
synthesis_generator, expression_generator = load_pretrained_model()
# Synthesize MIDI
output = synthesize_midi(synthesis_generator, expression_generator, midi_file)
# The synthesized audio
synthesized_audio = output['mix_audio']

Advance Usage

Here is an advance example to synthesize the ode_to_joy.mid, change the note expression controls, and adjust the
synthesis parameters:

import numpy as np
import tensorflow as tf
from midi_ddsp.utils.midi_synthesis_utils import synthesize_mono_midi, conditioning_df_to_audio
from midi_ddsp.utils.inference_utils import get_process_group
from midi_ddsp.midi_ddsp_synthesize import load_pretrained_model
from midi_ddsp.data_handling.instrument_name_utils import INST_NAME_TO_ID_DICT

# -----MIDI Synthesis-----
midi_file = 'ode_to_joy.mid'
# Load pre-trained model
synthesis_generator, expression_generator = load_pretrained_model()
# Synthesize with violin:
instrument_name = 'violin'
instrument_id = INST_NAME_TO_ID_DICT[instrument_name]
# Run model prediction
midi_audio, midi_control_params, midi_synth_params, conditioning_df = synthesize_mono_midi(synthesis_generator,
                                                                                           midi_file, instrument_id,

synthesized_audio = midi_audio  # The synthesized audio

# -----Adjust note expression controls and re-synthesize-----

# Make all notes weak vibrato:
conditioning_df_changed = conditioning_df.copy()
note_vibrato = conditioning_df_changed['vibrato_extend'].value
conditioning_df_changed['vibrato_extend'] = np.ones_like(conditioning_df['vibrato_extend'].values) * 0.1
# Re-synthesize
midi_audio_changed, midi_control_params_changed, midi_synth_params_changed = conditioning_df_to_audio(
  synthesis_generator, conditioning_df_changed, tf.constant([instrument_id]))

synthesized_audio_changed = midi_audio_changed  # The synthesized audio

# There are 6 note expression controls in conditioning_df that you could change:
# 'amplitude_mean', 'amplitude_std', 'vibrato_extend', 'brightness', 'attack_level', 'amplitudes_max_pos'.
# Please refer to https://colab.research.google.com/github/magenta/midi-ddsp/blob/main/midi_ddsp/colab/MIDI_DDSP_Demo.ipynb#scrollTo=XfPPrdPu5sSy for the effect of each control. 

# -----Adjust synthesis parameters and re-synthesize-----

# The original synthesis parameters:
f0_ori = midi_synth_params['f0_hz']
amps_ori = midi_synth_params['amplitudes']
noise_ori = midi_synth_params['noise_magnitudes']
hd_ori = midi_synth_params['harmonic_distribution']

# TODO: make your change of the synthesis parameters here:
f0_changed = f0_ori
amps_changed = amps_ori
noise_changed = noise_ori
hd_changed = hd_ori

# Resynthesis the audio using DDSP
processor_group = get_process_group(midi_synth_params['amplitudes'].shape[1], use_angular_cumsum=True)
midi_audio_changed = processor_group({'amplitudes': amps_changed,
                                      'harmonic_distribution': hd_changed,
                                      'noise_magnitudes': noise_changed,
                                      'f0_hz': f0_changed, },
midi_audio_changed = synthesis_generator.reverb_module(midi_audio_changed, reverb_number=instrument_id, training=False)

synthesized_audio_changed = midi_audio_changed  # The synthesized audio


View Github