anujpaude1/moodwave-music-tagging

MoodWave Music Tagging

A multilingual music-tagging dataset built for Hindi and Nepali repertoire, which existing music models handle poorly. 20,939 tracks, 1,763 hours.

This repository contains metadata only — no audio. Each row carries a YouTube ID (or a content hash for legacy entries) plus labels and provenance. See Getting the audio.

Status: work in progress. Collection is ongoing and genre labels currently cover the English portion only. See Current state before using this — the numbers below are live, not targets.

Current state

	Count
Total tracks	20,939
With a genre label	20,710
Unlabelled (awaiting the labeling pass)	229
Total audio	1,763 hours

By language

Language	Tracks
hi	10,202
ne	4,906
en	4,672
instrumental	335
pa	290
other	208
unlabelled	155
new	118
ur	40
bho	12
mai	1

By genre

Genre	Tier	Tracks
Pop	A	6,185
Ghazal	B	1,630
Bhajan/Devotional	B	1,384
Country	A	1,286
Adhunik	B	1,222
Hindustani Classical	B	1,157
Folk	A	1,124
Bhangra/Punjabi Folk	B	1,072
Nepali Lok Geet	B	994
Jazz	A	870
Lok Dohori	B	783
Hip-Hop	A	550
Reggae	A	481
Classical	A	457
Rock	A	450
Electronic	A	331
R&B	A	293
Tamang Selo	B	292
Sufi/Qawwali	B	89
Deuda	B	60

Why the taxonomy looks unusual

"Bollywood" is not a genre here. It describes who commissioned the music, not what it sounds like — filmi spans Hindustani classical, folk, rock, qawwali and EDM. A Hindi film song therefore carries language=hi, is_filmi=true and a real musical genre.

Encoding Bollywood as a genre teaches a model "Hindi language ⇒ Bollywood", which then mislabels every Hindi rock song. This follows Apple Music (Primary Bollywood + Secondary Soundtrack), Spotify (bollywood is a child of pop) and MusicBrainz (no bollywood genre at all, only filmi).

Genres

Tier A — language-agnostic sound: Pop, Rock, Hip-Hop, Electronic, Folk, Jazz, R&B, Classical, Country, Reggae

Tier B — regional traditions with a distinct acoustic signature: Hindustani Classical, Sufi/Qawwali, Bhajan/Devotional, Ghazal, Bhangra/Punjabi Folk, Lok Dohori, Tamang Selo, Deuda, Nepali Lok Geet, Adhunik

Blues folds into R&B and Metal into Rock. Classical means the Western art-music tradition; Indian classical is Hindustani Classical (which also absorbs Carnatic).

Two boundary rules

1. Newa Dapha and Gunla Bajan are Nepali Lok Geet, not Bhajan/Devotional. The Devotional class is defined by sound — harmonium + dholak/tabla + manjira — not by religious function. Newar traditions use dhimay/khin drums and asta-dhatu cymbals, and no harmonium.
2. genre=Deuda never implies language=ne. Deuda is performed in the Kumaon Division of Uttarakhand, India as well as in far-western Nepal.

Schema

Column	Type	Notes
track_id	string	yt_<11-char YouTube ID>, or lg_<11 hex> for legacy entries
youtube_id	string	Empty for legacy entries
audio_sha256	string	Content hash of the encoded audio; the dedup key
title	string	Original upload title, verbatim (may be Devanagari)
title_romanized	string	Song name with promotional furniture stripped
artist / artist_id	string	Populated only when confidently resolved (see below)
composer	string	Music director, for filmi
genre / genre_tier	string	20 classes, tier A or B. Empty = unlabelled
tradition	string	Sub-label: salaijo, jhyaure, teej, newa_dapha, …
language	string	hi ne en pa bho new mai ur other instrumental
origin_region	string	NP IN PK other
is_filmi / film_title / film_industry	bool / string	Film provenance
release_year / decade	int16 / string	-1 when unknown
mood / mood_source	string	6 moods
vibes	list<string>	12 vibes
vocal_gender	string	male female duet chorus instrumental
is_live is_instrumental is_remix is_cover	bool
duration_s	float32
source_channel_id / source_playlist_id	string	Provenance
label_source / label_confidence	string / float32	Which signals agreed
taxonomy_version	string	Must match taxonomy/genres.json
split	string	train / unlabelled

A note on artist

artist is populated only when confidently resolved, and is otherwise empty. Both title orderings occur on the source channels:

Nepathya - Tadhikha                     <- Artist - Track
Hariyo Danda Mathi - Dharma Raj Thapa   <- Track - Artist

Nothing in the string settles which is which, so the split is only made when there is real evidence (a known single-artist channel, or comma-separated names on exactly one side). title_romanized always carries the full cleaned name.

Usage

from datasets import load_dataset

ds = load_dataset("anujpaude1/moodwave-music-tagging", split="train")
print(ds[0]["title_romanized"], "|", ds[0]["language"], "|", ds[0]["genre"])

# Labelled subset only
labelled = ds.filter(lambda r: r["genre"] != "")

# Nepali folk
dohori = ds.filter(lambda r: r["genre"] == "Lok Dohori")

Query without downloading, via the SQL Console or DuckDB:

SELECT language, genre, count(*) AS n
FROM 'hf://datasets/anujpaude1/moodwave-music-tagging/data/manifest/train-*.parquet'
WHERE genre != ''
GROUP BY 1, 2 ORDER BY n DESC;

Windows users: pin datasets<4.0. Version 4.0 moved audio decoding to torchcodec, which has no Windows build. The metadata in this repo is unaffected, but the audio repo below is.

Getting the audio

No audio is distributed here. To reconstruct it, resolve youtube_id yourself:

yt-dlp --js-runtimes node -f bestaudio \
       --extract-audio --audio-format best \
       -o "yt_%(id)s.%(ext)s" \
       --sleep-requests 3 --sleep-interval 6 --max-sleep-interval 15 \
       "https://www.youtube.com/watch?v=<youtube_id>"

The reference pipeline transcodes to Opus 48k mono 24 kHz full-length (not clipped). 24 kHz rather than 16 kHz so that MERT and CLAP remain usable; discogs-effnet resamples to 16 kHz and reads nothing above 8 kHz, so the encoding is transparent to it.

Legacy rows (lg_ prefix) have no YouTube ID — they came from an earlier collection and are identified by content hash only.

How the data was collected

1. Enumerate official label and artist channels with yt-dlp --flat-playlist.
2. Filter — reject anything outside 60–900 s, plus compilations and non-music. Label channels are not song channels: they also publish comedy series, trailers, teasers, interviews, BTS and episode promos, all of which sit inside the duration window. Duration alone let 19% non-music through on the first pass.
3. Download and transcode the full song, discarding the source immediately.
4. Label — retrieval first (iTunes, Last.fm, MusicBrainz, channel provenance), with an LLM used only to reconcile that evidence onto the fixed taxonomy. A label is accepted only on agreement between 2+ independent sources. (Not yet run for Hindi/Nepali.)

English genre labels come from a hand-curated ground-truth set of an earlier collection, joined by normalised filename at a 95% match rate.

Limitations

• Labels are incomplete. Genre currently covers the English portion only.
• No artist-disjoint splits yet. The corpus is built from single-artist playlists, so a random split leaks artist identity and inflates accuracy (Flexer 2007: 75.7% → 58.5% under an artist filter). Do not evaluate on a random split.
• Sampling bias toward what labels upload and promote; the Hindi catalogue skews heavily post-2000.
• Mood labels are weak. Audio-only Hindi mood classification tops out around 58% F-measure in the literature; lyrics are needed to do better.
• Romantic is a lyrical category with no position in valence–arousal space and will collapse into Calm and Sad under an audio-only model. It also appears on the vibe axis.

Licence

Annotations: non-commercial research use (see LICENSE.md). The underlying compositions and recordings remain the property of their rights holders. This repository distributes no audio.

For takedown requests, open a discussion on this repository.