ProjectAquarius/colour-regions

Colour Regions

Colour Regions is an open corpus of Chinese folk song for geographic and ethnomusicological research. It groups 1692 recordings (91.2 h) into the 15 colour regions (色彩区) of Chinese folk-song scholarship: 11 Han regions and 4 minority song areas.

This dataset contains no audio, and no audio file is ever uploaded: only tables, MIDI, MusicXML, note tables and F0 pitch tracks (pitch values in Hz, not sound). Each recording is a link to its source (1113 YouTube, 571 Bilibili and 8 Europeana/CREM items). The dataset gives the curation metadata and the data derived from the audio: note transcriptions (MIDI and MusicXML) and pitch tracks (F0). The 8 CREM archive items are "In Copyright, research use only": for them the dataset gives links and metadata only, without derived data.

Status: pre-release, private. Authors, citation and the accompanying paper (TISMIR, special collection "Open Music Data for Music Processing Research") will be added.

What is in it

Path	Content
data/recordings.parquet	One row per recording: link, region, place, ethnic group, song, genre, performance type, provenance, song labels, quality flags, and paths to the derived files.
data/song_labels.parquet	One row per song (1497 songs): song class, genre, composer, lyricist, arranger, period of composition, with the evidence, sources 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), 62 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, to check that a re-downloaded file is the same recording.
data/link_status.parquet	The latest check of whether each link still works.
midi/performance/	The transcription as sung (performance view): one MIDI per recording, the vocal melody.
musicxml/score/	The transcription as notated (score view): one MusicXML per recording, notated pitch and note boundaries. The metre is nominal: free-rhythm songs (散板, 长调) get a beat grid 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: a MIDI (performance view) and a MusicXML (score view). Instrumental items (no singing in the separated voice) and the CREM links-only items have no transcription files. All paths in recordings.parquet are relative to the dataset root, for example midi/performance/Northwest_Plateau/<id>.mid.

The two transcription views

• Performance view: majority vote, note by note, over 13 transcriptions of each recording: pretrained GAME (6 runs on the separated voice, 3 on the original mix) and SOME (2 checkpoints × voice and mix). Notes are split where the F0 settles on another semitone, and the sung pitch mixes the fused pitch with the note's median F0.
• Score view: GAME fine-tuned on score-annotated Chinese singing (M4Singer and GTSinger, every benchmark song held out), 3 runs on the separated voice, majority vote, onsets aligned to voicing onsets.

Note accuracy on the benchmark (pitch within ±50 cents, onset within ±150 ms, after removing each subset's constant onset bias):

	vocadito (sung-pitch reference)	GTSinger (score reference)	M4Singer (score reference)
performance view	0.81	0.76	0.72
score view	0.78	0.82	0.74
earlier transcription (3-run GAME vote, until 2026-10)	0.72	0.75	0.67
second human annotator	0.79	–	–

Recordings whose separated voice holds no singing (instrumental items) have no transcription files.

How the data was made

1. Collection. Three rounds of curated web search:
   • Round 1: 600 YouTube recordings, 40 per region.
   • Round 2: 697 recordings from new channels, mainly Bilibili, with a channel cap and provenance rules.
   • Round 3: 358 recordings from the Musical Map of China series (中国音乐地图), whose descriptions name the singers and places.
   • Round 4: 22 recordings from a Northeast Plain playlist (mostly Musical Map of China items: 二人转 pieces, 秧歌, suona).
   • Round 5: 15 recordings from a Beijing playlist (Peking opera excerpts, percussion ensembles, orchestral and instrumental pieces; most are not folk song). Every link comes from a logged search; none was written from memory.
2. Region labels and provenance. Each recording has a provenance tier:
   • A1: national heritage bearer;
   • A2: named singer with a stated county;
   • B: field or locality-attested recording;
   • C: region from title or genre only.
3. Transcription.
   • htdemucs separates the voice from the accompaniment.
   • Performance view: 13 transcriptions (GAME and SOME, on the separated voice and the mix), combined by note-level voting and refined with the F0 track. Score view: GAME fine-tuned on score-annotated singing, 3 runs, voted.
   • RMVPE and PESTO compute F0 on the separated voice.
4. Song labels. Research agents labelled each song from sources, following a written protocol. The labels were cross-validated against an earlier blind dating (94% agreement on the song class) and against the Anthology, and then checked by an independent verifier. A second, independent review covered every song. The column song_label_check says what that review found: verified (checked against a source), reviewed (consistent with all the evidence, no source read), corrected (changed after review) or disputed (an open question).

Main columns of recordings

Column	Meaning
item_id, platform, url	Source identifier and link
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 source and 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
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)
subset	core (default analysis subset, 1570), folk_style_composed (songs composed in a folk idiom, 32), excluded (label errors, arrangements, composed songs, speech)
group_id	Connected component of "same channel" ∪ "same singer". Keep a group on one side of any train/test split.
transcription_model, n_notes, vocal_db, game_rosvot_agreement	Quality indicators of the earlier transcription
midi_performance, musicxml_score, f0_*	Paths to the derived files

Getting the audio

The audio is third-party and is not redistributed. Download each recording from its url under the platform's terms (for example with yt-dlp), then check it against fingerprints (duration within ±2 s, Chromaprint bit error rate ≤ 0.25). Some links will stop working over time: see link_status. The derived data stays usable without the audio.

Evaluation

Recordings from one channel or one singer share recording conditions and voice. A model can learn those instead of the region. Use group_id to keep each group on one side of a split, and keep songs (song_key) disjoint. The Musical Map of China series (flag series:musical_map_of_china, 539 recordings) forms most of the largest group; it spans all 15 regions, so its channel is not a shortcut to the region label, but report results with and without it.

Limitations

• Labels: most song labels are low confidence: anonymous traditional songs labelled from the genre and the title. Region labels of tier C items rest on the title or genre only.
• Balance: regions are unequal in size (73–202 recordings). The Musical Map of China series adds many minority-region items from one studio production.
• Transcriptions: they are automatic. The benchmark is studio and solo singing, not folk recordings; in-domain checks show agreement with Anthology scores of the same song, but individual notes are often wrong, especially in heterophonic, ornamented or free-rhythm singing. No human-checked folk ground truth exists yet.
• Platform bias: the sources are online videos, which over-represent staged and broadcast performances.

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. Video titles are third-party text, kept as published. 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.