ProjectAquarius/colour-regions

Colour Regions

Pre-release: version v0. The tables, file layout and labels may still change before the first stable release.

Colour Regions is an open corpus of Chinese folk song for geographic and ethnomusicological research. It groups 1602 recordings (85.8 h) into the 15 colour regions (色彩区) of Chinese folk-song scholarship: 11 Han regions and 4 minority song areas. For each recording it gives curation metadata, song labels, a note transcription (MIDI and MusicXML) and F0 pitch tracks.

What is in it

Path	Content
data/recordings.parquet	One row per recording: region, place, ethnic group, song, genre, performance type, provenance, song labels, quality indicators, and paths to the files below.
data/song_labels.parquet	One row per song (1441 songs): song class, genre, composer, lyricist, arranger, period of composition, with the evidence and confidence of each label.
data/anthology_overlap.parquet	Songs of this corpus that are also printed in the Anthology of Chinese Folk Songs (《中国民间歌曲集成》, OMR edition): 61 songs matched.
data/scores.parquet, data/score_matches.parquet	Anthology and Essen scores mapped to the same regions (statistics only), and title matches between recordings and scores.
data/fingerprints.parquet	Chromaprint fingerprints of the first 120 s of each recording.
data/link_status.parquet	The latest availability check of each recording.
midi/performance/	The transcription as sung (performance view): one MIDI file per recording, the vocal melody.
musicxml/score/	The transcription as notated (score view): one MusicXML file per recording, with notated pitch and note boundaries. The metre is nominal: free-rhythm songs (散板, 长调) get a beat grid that they do not have. Use the MIDI for timing.
f0/rmvpe/, f0/pesto/	F0 of the separated voice, 10 ms hop, in Hz (0 = unvoiced). PESTO files also hold a confidence.

Each recording has one transcription, given in two views. 1550 recordings have a transcription. 52 have none: 10 have no singing found in the separated voice (they keep their F0), 34 are not transcribed yet, and 8 are metadata-only by licence. All paths in recordings.parquet are relative to the dataset root, for example midi/performance/Northwest_Plateau/<id>.mid.

Performance view and score view

The two views answer different questions about the same singing.

• Performance view (midi/performance/): what the singer actually sings. Each note keeps its sung timing, and slides and ornaments split into the notes the voice reaches. Use it to study intonation, ornament and timing.
• Score view (musicxml/score/): what a notator would write down. The model learned from score-annotated singing, so it gives the notated pitch and merges ornaments into the main note, as in a printed score. Use it to compare with scores such as the Anthology, or to read the melody.

The two views come from different transcription models, so they can differ in their number of notes as well as in pitch. In both, MIDI pitches are whole semitones.

How to load

from datasets import load_dataset
from huggingface_hub import hf_hub_download

rec = load_dataset("ProjectAquarius/colour-regions", "recordings", split="train").to_pandas()
row = rec[rec.midi_performance.notna()].iloc[0]
midi = hf_hub_download("ProjectAquarius/colour-regions", row.midi_performance, repo_type="dataset")

Other tables load the same way, with the config names song_labels, anthology_overlap, scores, score_matches, fingerprints and link_status.

Main columns of recordings

Column	Meaning
item_id	Recording identifier
region, region_en, region_type	Colour region (Chinese, English), Han or minority
region_curator	The curator's region, where a later decision moved the recording
province, county_or_area, lon, lat, geo_level	Stated place and its location
ethnic_group, song_name, song_key, genre, performance_type	As stated by the curators
alt_titles	Verified alternative titles ("; "-separated): a title is listed only when a published score printed under it matched the song's recordings by melody
singer, singer_id	Names are given only for national heritage bearers; singer_id is a pseudonymous ID
provenance_tier, provenance_evidence	Strength of the region attestation, and the evidence: A1 national heritage bearer; A2 named singer with a stated county; B field or locality-attested recording; C region from title or genre only
feature_flags	List of semantic features, e.g. series:musical_map_of_china, accompanied, instrument:马头琴, multiple_singers
song_class, song_genre, composer, composition_period, …	Song labels (see song_labels for the evidence)
song_label_check	Review status of the song label: verified (checked against a published source), reviewed (consistent with all the evidence), corrected (changed after review), disputed (an open question) or unchecked (not reviewed yet; 25 recordings)
subset	core (default analysis subset, 1570) or folk_style_composed (songs composed in a folk idiom, 32)
group_id	Connected component of "same channel" ∪ "same singer". Keep a group on one side of any train/test split.
n_notes, notes_per_s	Size of the shipped transcription (performance view)
vocal_db, sung_share, music_share	Level of the separated voice, and the share of 10 s windows classified as singing and as music
midi_performance, musicxml_score, f0_rmvpe, f0_pesto	Paths to the files

Licence and removal

The metadata and derived data are released under CC BY-NC-SA 4.0. Data derived from the Anthology of Chinese Folk Songs OMR edition (CC BY-NC-SA 4.0) keeps that licence. Personal names appear only for national heritage bearers, a public role.

To ask for an item to be removed, open a discussion on this page.

Versions

• v0 (pre-release): 1602 recordings; one transcription per recording in two views (performance MIDI, score MusicXML); F0; song labels with a second review; verified alternative titles.