Cite us

This is the repo of all codes used in our paper [...].

Setup

Install pyenv and poetry. You can also setup your environment with other tools, if the dependencies listed in pyproject.toml are installed.

Environment

The following is an example of setup for venv.

1. cd <repo>
2. git clone <this-repo-url>
3. pyenv install --patch 3.6.9 < python_alignment.patch; pyenv local 3.6.9 or just use python 3.6.9; note that the install command is using this patch
4. python -m venv .venv; source .venv/bin/activate.fish
5. python -m pip install -U pip
6. python -m pip install -r requirements.txt
7. pip install git+https://github.com/CPJKU/madmom -c constraints.txt (this is needed because magenta has an exact dependency for an old mido version in its requirements.txt)
8. python setup.py build_ext --inplace

Then, you'll need vienna_corpus, SMD and Maestro datasets from asmd package:

python -m asmd.install

Chose excerpts

1. Download our pretrained vienna model on Maestro and put it in your working dir from our mega
2. Train our proposed model or download the pretrained ones from our mega:
3. You will need the template matrix provided in this repo. To rebuild it run python -m perceptual.make_template. You will need the synthesized scale and the corresponding midi in the scales and audio folder. You can download them from our mega
   1. python -m perceptual.proposed create_mini_specs to create the dataset of mini-specs or download it from our mega.
   2. dataset size: 474.429 notes (831 batches in test, 178 in train))
   3. python -m perceptual.proposed train to train our model for velocity estimation and test it. I obtained the following absolute error (avg, std) on the test set: 15.11, 10.94 (251 epochs)
   4. redo everything with vienna model (use --vienna for create_mini_specs and train)
4. Run python -m perceptual.excerpt_search

This will analyze vienna_corpus in search of excerpts, will transcribe the original performances and will create a new directory audio with all extracted excerpts audio files and a directory to_be_synthesized with all midi files that you have to synthesize and put in audio

Chose vsts

1. Synthesize the chosen excerpts with vsts or download our synthesized midis from our mega; extract them in the audio directory. You should have a directory for each vst in audio and for each vst you should have 5 different audio. In the root of audio you should also have the original recordings.
2. Install sox in your path for post-processing to add reverb or run with --no-postprocess
3. Analyze chosen excerpts: python -m perceptual.chose_vst

This will copy the excerpts relative to the chosen vsts to the folder excerpts.

Chosen vsts: - q0: ./audio/salamander - q1: ['./audio/pianoteq1', './audio/salamander-norm-20_reverb_50_norm'] - q2: ./audio/pianoteq1-norm-20_reverb_100_norm

Build tests

Set up your server (Python or PHP) and download WAET.

1. place the directory excerpts in the root of WAET
2. place the directory reveal.js-3.9.2 into the root of WAET
3. place the file index.html in the root of WAET (if you want, you can regenerate the index.html by running pandoc --to revealjs -V revealjs-url=reveal.js-3.9.2 --output index.html --standalone index.md)
4. place the file listening_test.xml in [WAET root]/tests/pool.xml
5. place the file core.css in [WAET root]/css/core.css

You should be able to access your test at /test.html?url=php/pool.php. More info in the WAET wiki

index.html contains the instructions for the test, so that you can distribute the url to the root of WAET to your partecipants.

Answer analysis

To plot tests you should use streamlit run perceptual_app.py, which also prints correlations with the objective measure of your choice.

The test answers that we collected are available in the repo.

Before of running you should change the settings according to your system: open the script and change the initial global variables:

1. PATH is the path to the saves dir of WAET
2. DISCARD_BEFORE_THAN defines a date before of which the answers whould be discarded; this is useful for removing debug answers
3. MAP_VALUES defines the mapping for creating the control groups according to the answer of the users

Also note that all answers in which the users listened to for less than 5 seconds or didn't move the cursor are completely discarded. This is hard-coded in final section of the script.

At each run, violin plots are created for each control group and each method. One plot is created for each question type and excerpt or for each question type if average option is used. Under each plot, there are the p-values computed for each combination of groups or methods. The error margins and correlations are shown too.

Supplementary materials show some of the plots that can be generated.

Linear regression

To compute the linear regressions of the perceptual values, you should run python perceptual.eval_regression. It will plot the regression predictions for various model and weights for the case with and without MFCC features. Than, it will also plots the weights with only the selected features.

If you want, you can test the selected features by using our_eval as option to the subjective_eval script.

Alignment

1. Install fluidsynth and download SalamanderGrandPianoV3 soundfont in sf2 format from our mega folder and put it in your working dir
2. run python -m perceptual.alignment.dtw_tuning to check the FastDTW tuning in midi2midi over MusicNet solo piano songs
3. run python -m perceptual.alignment.align amt to perform our amt-based alignment over SMD dataset with the best parameters found in the previous step
4. run python -m perceptual.alignment.align ewert to perform our baseline alignment over SMD dataset
5. run python -m perceptual.alignment.analysis results/ewert.csv results/amt.csv to plot the results of alignment

Credits

1. Federico Simonetta