/adt_with_a2md

This is the material for paper "IMPROVING AUTOMATIC DRUM TRANSCRIPTION USING LARGE-SCALE AUDIO-TO-MIDI ALIGNED DATA"

Primary LanguageJupyter NotebookGNU General Public License v3.0GPL-3.0

Supplementary material for paper:

“IMPROVING AUTOMATIC DRUM TRANSCRIPTION USING LARGE-SCALE AUDIO-TO-MIDI ALIGNED DATA”

*** IMPORTANT CORRECTION !! ***

The numbers reported in the third column of Table 2 (B-CNN, 3-fold) are presented improperly. They should be 0.75, 0.72, and 0.60 (marked in red below) for the three public datasets. The results in Table 2 still verify the effectiveness of the proposed model despite the correction.

This error was due to the inconsistent configuration of the 3-fold cross validation process. In the previous work [1], the evaluation results based on two different 3-fold strategies were reported on their complementary website (http://ifs.tuwien.ac.at/~vogl/dafx2018/):
[i] sepearate 3-fold -- perform 3-fold cross validation on each of the public datasets (ENST/MDB-Drums/RBMA13)
[ii] joint 3-fold -- combine all public datasets and perform 3-fold cross validation jointly.
In our submitted manuscript, we included the results of CNN/CRNN (3-fold) under scenario [ii] in Table 2 for comparison. However, we incorrectly reported the results of B-CNN (3-fold) under scenario [i], which led to an inconsistent comparison with the prior work. To avoid confusions and ensure the compatibility of the results, we have updated the table (as shown above) and will correct the numbers in the manuscript accordingly. Note that the updated results in Table 2 are coherent with our observations and do not change our conclusions in the submitted manuscript.


[1] Richard Vogl, Gerhard Widmer, and Peter Knees, “Towards multi-instrument drum transcription,” in Proceedings of International Conference on Digital Audio Effects (DAFx), 2018.

Accompanying materials:

  • Drum transcription tool: The implemented ADT system is integrated in python package Omnizart.

  • Penalty search: A Python script to demo the process of penalty value retrieval for a given audio pair. Please check file penalty_search.ipynb for more details.

  • A2MD dataset: 1565 internet downloaded audio tracks and their aligned MIDI files are provided in a single zip file. The audio/MIDI files are categorized into seven groups based on the retrieved penalty values. A2MD download link:
    https://drive.google.com/uc?export=download&id=1ZRbLz7aaJibd9F121LLn_yKPmgtYUqHu

  • Song demo: We provide three demo songs for model comparison. In the demo songs, the original drum track is replaced with re-syntheszed drum track using transcription results from different models. It should be noted that only three basic drum instruments (i.e., kick drum, snare drum, hihat) are used in the the re-synthesized track. Please check directory demo_songs for audio tracks.