Context-aware Automatic Music Transcription

Supplementary information are available on the MIA website.

Setup

1. Enter the root git directory
2. Install poetry
3. Install pyenv
4. Install python 3.8.6: pyenv install 3.8.6
5. Create a new venv with poetry and install the dependencies: poetry install --noroot
6. Start a new shell in this venv: poetry shell
7. Check that the version is correct with python --version

Config

1. Open mpc2c/settings.py and check the variables in section 'PATH', especially VELOCITY_DATA_PATH, PEDALING_DATA_PATH, and RESYNTHESIS_DATA_PATH. Set them to meaningful paths for your system.
2. Make sure that you have jackd installed
3. Run poetry run python -m pycarla.carla -d to download Carla host

4. Prepare your Carla presets by using poetry run python -m pycarla.carla -r and put them in the mpc2c.settings.CARLA_PROJ directory; the included ones need:

   • Pianoteq 6 STAGE LV2 plugin installed and available to Carla (e.g. in ~/.lv2/ or /usr/lib/lv2/)
   • Calf Reverb LV2 plugin installed and available to Carla (e.g. in ~/.lv2/ or /usr/lib/lv2/)
   • SalamanderGrandPianoV3Retuned SFZ version installed in /opt/soundfonts/SalamanderGrandPianoV3+20161209_48khz24bit/SalamanderGrandPianoV3Retuned.sfz

Datasets

1. Install 'Maestro' dataset from asmd: python -m mpc2c.asmd.asmd.install
2. Install Carla with python -m mpc2c.pycarla.pycarla.carla -d
3. Install jackd in your path
4. Prepare the new dataset with the resynthesized parts: python run.py -d
5. If the process stops, re-launch it (it will skip the already synthesized songs)
6. If it fails, it's probably because Carla crashed; just stop (CTRL+C) and restart (it will skip already synthesized songs); try to killall -9 jackd before restarting the process.
7. After having synthesized , you can do a full check that everything has correctly been synthesized by rm asmd_resynth.txt and relaunching the process python run.py -d
8. The datasets were split using PCA and retaining 0.89, 0.93, 0.91 of total variance for train, validation and test set respectively.

Command-line options

For all the available commands, using -v allows to extract velocity data, while using -p refers to pedaling.

Other options are described below.

1. Preprocess

1. Create the MIDI file for the template, synthesize and compute the template: python run.py -sc
2. Apply NMF and extract notes for [velocity|pedaling] estimation: python run.py [-v|-p] -r

2. Training the models

1. Evaluate [velocity|pedaling] configurations using: python run.py [-v|-p] -sk.
2. Option -cm cleans MLFlow runs, use it if the previous command fails for some reason, because the final evaluation is based on MLFlow

3. Evaluation

Run python run.py -e [-v|-p] to evaluate the average L1 error in each configuration

---

The previous steps can be done in a single command: python run.py [-v|-p] -sc -r -sk -e

Notes

We used 6 different artificial contexts generated from Pianoteq LV2 plugin. The following table shows the differences among the 6 contexts:

Context	Velocity Map	Reverb	Instrument
pianoteq0	Linear	Jazz Studio	Steinway B Prelude
pianoteq1	Logarithmic	Jazz Studio	Steinway B Prelude
pianoteq2	Logarithmic	Cathedral	Steinway B Prelude
pianoteq3	Linear	Jazz Studio	Grotrian Cabaret
pianoteq4	Logarithmic	Jazz Studio	Grotrian Cabaret
pianoteq5	Logarithmic	Cathedral	Grotrian Cabaret

The Carla project files available in the repo allow to see each individual parameter of the contexts.

Optional

Fully train models with: -t option. After each training, you will find a few checkpoint files in the relative directory directory; find the most recent one with ls -t | head -1. The hyperparameters used are not the best ones (you can change them in settings.py) and models are trained using multiple performers. Context specificity can be turned on with -cs option.

Credits

1. Federico Simonetta