Aligned Scores and Performances (ASAP) dataset

ASAP is a dataset of aligned musical scores (both MIDI and MusicXML) and performances (audio and MIDI), all with downbeat, beat, time signature, and key signature annotations.

This repository contains all of the annotations, as well as all of the MIDI and MusicXML files. To get the audio files, follow the instructions below.

Getting the Audio:

The audio files are not distributed in this repository. To obtain them:

• Install dependencies:
  • python 3
  • librosa
  • pandas
  • numpy
• If you use conda, you can install the dependencies with:
  • conda env create -f environment.yml
• download the Maestro dataset v2.0.0 https://magenta.tensorflow.org/datasets/maestro
• unzip the data
• run python initialize_dataset.py -m [maestro location]
• The maestro directory and zip file can now be safely deleted

The script has been tested in Windows, Linux and Mac OS with python 3.6, and the libraries librosa v0.7.2 and pandas v1.0.3.

Citing

If you use this dataset in any research, please cite the relevant paper:

@inproceedings{asap-dataset,
  title={{ASAP}: a dataset of aligned scores and performances for piano transcription},
  author={Foscarin, Francesco and McLeod, Andrew and Rigaux, Philippe and Jacquemard, Florent and Sakai, Masahiko},
  booktitle={International Society for Music Information Retrieval Conference {(ISMIR)}},
  year={2020},
  pages={534--541}
}

Dataset Content

ASAP contains 236 distinct musical scores and 1067 performances of Western classical piano music from 15 different composers (see Table below for a breakdown).

Scores and performances are distributed in a folder system structured as composer/subgroup/piece where the piece directory contains the XML and MIDI score, plus all of the performances of a specific piece. subgroup contains additional hierarchies (e.g., Bach contains subgroups Fugues and Preludes).

A metadata CSV file is available in the main folder with information about file correspondances, and the title and composer of each performance.

Annotations are given in tab-separated-values (TSV) files in the same folder as each performances, named as {basename(file)}_annotations.txt. These TSV files can be read by Audacity (File->Import->Labels) to view the annotations on a corresponding audio perfomance. Paired MIDI and audio performances correspond to the same annotation file since they are exactly aligned.

A single json in the main folder also contains all of the annotation information for every file in ASAP.

It is possible that 2 versions of the same piece are in different folders if they refer to slighlty different scores (e.g. different repetitions in the performance). Even in this case, the composer and title in the metadata table will be the same for both performances. For the applications where unique pieces are needed (e.g., to create a training/test dataset with not overlapping) look for the unique couple (title,composer). Note that this is not the same as considering all the unique xml-score names, since in order to deal with different repetitions in performances, 2 different xml-scores may refer to the same piece (e.g., "Beethoven/Piano_sonatas/17_1/xml_score.musicxml" and "Beethoven/Piano_sonatas/17_1_no_repeat/xml_score.musicxml").

Metadata table

Each row in metadata.csv file contains the following information:

• composer: a common name for the composer of a piece
• title: a title that (along with composer) uniquely identifies each piece in the dataset
• folder: the path of the directory containing all scores and performances of the piece
• xml_score: the path of the MusicXML score of a piece
• midi_score: the path of the score in MIDI format
• midi_performance : the path of the MIDI performance
• performance_anotations : the path of the TSV annotations on the performances (audio and MIDI)
• midi_score_annotations : the path of the TSV annotations of the MIDI score
• maestro_midi_performance : the path of the MIDI performance in the MAESTRO dataset (if it exists)
• maestro_audio_performance : the path of the audio performance in the MAESTRO dataset (if it exists)
• start: If not blank, specifies a time in seconds where the original MAESTRO performance (audio and MIDI) has been cut to match the annotations. The ASAP performance at time 0 will match the original MAESTRO performance at this time
• end: The same as start, but for the end of the original MAESTRO performance. The end of the ASAP performance is at this time in the original MAESTRO performance
• audio_performance: the path of the (properly cut) audio file in the ASAP dataset (if one exists)

Annotation json

In asap_annotations.json the first group of keys are the path of the MIDI performance files in ASAP. Every performance has the following keys:

• performance_beats : a list of beat positions in seconds
• performance_downbeats : a list of downbeat positions in seconds
• performance_beats_type : a dictionary where the annotation time is the key and the value is either "db"(downbeat), "b" (beat) or "bR" (in cases where standard notation rules are not followed---e.g. rubato or pitckup measures in the middle of the score---and we cannot determine the exact beat position)
• perf_time_signatures : a dictionary where the annotation time is the key and the value is a couple [time_signature_string,number_of_beats]. number_of_beats is an integer specifying the number of beats per measure. For compound meters, there is not a universal agreement on the number of beats; here we refer to https://dictionary.onmusic.org/appendix/topics/meters, and annotate e.g., 6/8 with 2 beats per measure.}
• perf_key_signatures : a dictionary where the annotation time is the key and the value is a couple [key_signature_number,number_of_sharps]. The first value is an integer in [0,11] marking the tonic of the key signature (if it is major) with C = 0. The second is an integer in [-11,11] where negative numbers mark the number of flats and positive numbers mark the number of sharps. Note that we only annotate the key \emph{signature}, and not the key, since modes are not annotated in scores, so we cannot distinguish between, e.g., a major key and it's relative minor.
• midi_score_beats : same as perf_beats, but on the MIDI score
• midi_score_downbeats : same as perf_downbeats, but on the MIDI score
• midi_score_beats_type : same as perf_beats_type, but on the MIDI score
• midi_score_time_signatures : same as perf_time_signatures, but on the MIDI score
• midi_score_key_signatures : same as perf_key_signatures, but on the MIDI score
• downbeats_score_map : a list of measure numbers in the score (starting from 0, including any pickup measures) corresponding to each downbeat in the MIDI score. For example, if the first number in this list is 1, that means that the first annotated downbeat corresponds to measure number 1 in the score. In case multiple score measures correspond to the same downbeat (e.g., for a measure that is split in the score for some reason), the values in the list are a string in the form "number1-number2-number3-..."
• score_and_performance_aligned : True if the MIDI score and the performance have the same number of annotations. If this is false, the performance can still be used for beat tracking tasks, but not for full transcription. It is False for only 29 performances where some beats are skipped by the performer due to a mistake or an incomplete performance.

TSV annotation

For each performance and MIDI score we provide a tab-separated value (TSV) file of the position (in seconds) of each beat, downbeat, time signature change, and key signature change. These files can be read by Audacity to view the annotations on a corresponding audio perfomance. Each line in the TSV file begins with 2 identical columns containing the time in seconds corresponding to the annotation, followed by a 3rd column containing the label:

• Beats and downbeats: annotated with either b (for beats), db (for downbeats) or bR (in cases where standard notation rules are not followed---e.g. rubato or pitckup measures in the middle of the score---and we cannot determine the exact beat position)
• time signature changes: the time signature as a string (e.g., 4/4).
• key changes (k in [-11,11]): the number of accidentals in the key signature, where 0 is none, positive numbers count sharps and negative numbers count flats.

Usage examples

import pandas as pd
from pathlib import Path
import json
BASE_PATH= "."

#get a list of performances such as there are not 2 performances of the same piece
df = pd.read_csv(Path(BASE_PATH,"metadata.csv"))
unique_df = df.drop_duplicates(subset=["title","composer"])
unique_performance_list = unique_df["midi_performance"].tolist()

#get the downbeat_list of a performance of Bach Fugue_bwv_848
midi_path = df.loc[df.title=="Fugue_bwv_848","midi_performance"].iloc[0]
with open(Path(BASE_PATH,'asap_annotations.json')) as json_file:
    json_data = json.load(json_file)
db_list = json_data[midi_path]["performance_downbeats"]

#same task, but using the TSV file
annotation_path = df.loc[df.title=="Fugue_bwv_848","performance_annotations"].iloc[0]
ann_df = pd.read_csv(Path(BASE_PATH,annotation_path),header=None, names=["time","time2","type"],sep='\t')
db_list = [row["time"] for i,row in ann_df.iterrows() if row["type"].split(",")[0]=="db"]

#get all pieces with time signature changes
with open(Path(BASE_PATH,'asap_annotations.json')) as json_file:
    json_data = json.load(json_file)
tsc_pieces = [p for p in json_data.keys() if len(json_data[p]["perf_time_signatures"])>1 ]

Limits of the dataset

• The scores were written by non-professionals, and although we manually corrected them to filter out most of the incorrect notation, they still present some problems. Our suggestion is to use the corrected annotations provided in the TSV and in the JSON files rather than extracting them again from the score.

• In certain cases (e.g., pickup measures in the middle of the score, rubato measures, complex embellishments) it is not possible to estabilish the position of the beat from the score. In our annotations we choose to mark those beats as "bR".

• 31 performances are not aligned with the score. The cause is an incomplete performance or the player missing some beats due to a mistake. Altough not good for AMT and score production, those performances were manually checked and are still usable for beat/downbeat tracking. This information is in the json file.

• We created this dataset to target MIR tasks such as beat/downbeat tracking, key signature estimation, time signature estimation. We haven't explored other tasks such as voice separation, beaming/tuplet creation, and expressive performance rendering. It is surely possible to perform them with the information from the xml-scores and our alignment, but we haven't tested if the data is of a sufficient quality for those tasks.

License

The dataset is made available under a Creative Commons Attribution Non-Commercial Share-Alike 4.0 (CC BY-NC-SA 4.0) license.

Versions

• The current version of ASAP (v1.1) has some corrections, as enumerated in the change notes on the tag v1.1.
• The initial version of ASAP (as it was released at ISMIR 2020) is available as v1.0.