S16k — Multi-label Music Emotion Recognition from NetEase Cloud Music

169,148 songs with multi-label emotion annotations, the 46,613 user-curated playlists the annotations are derived from, pre-extracted audio features, and a class-balanced subset.

Reference paper: https://doi.org/10.1007/s00530-025-01701-z

How the labels are defined

Every playlist on NetEase Cloud Music (网易云音乐) carries editorial tags such as 流行, 治愈, 夜晚. A song inherits the tags of every playlist that contains it, and the value stored is a count: emo_healing = 3 means three playlists holding this song are tagged 治愈.

These are platform and community tags, not controlled psychological ratings. Treat them accordingly when comparing against annotation-based corpora such as DEAM or PMEmo.

Files

File	Rows	Size	Content
songs.parquet	169,148	52 MB	Main table: metadata, 74 tag columns, emotion labels
songs_9822.parquet	9,822	5 MB	Class-balanced subset, same schema plus unique_id
playlists.parquet	46,613	10 MB	Source playlists with tags and popularity statistics
song_playlist.parquet	477,216	2.4 MB	Song ↔ playlist relation, long format
tag_taxonomy.parquet / .csv	74	<1 MB	Tag column ↔ English ↔ Chinese ↔ group mapping
songs_169148_vggish.npz	169,097	2.7 GB	VGGish embeddings
songs_9822_vggish.npz	9,822	0.16 GB	VGGish embeddings for the balanced subset
spectrograms_169148.npz	169,148	17.3 GB	Log-mel spectrograms in dB

Audio features are computed from the middle 30 seconds of each track. Raw audio is not redistributed.

Quick start

import pandas as pd

songs = pd.read_parquet("songs.parquet")
tax   = pd.read_parquet("tag_taxonomy.parquet")

emo_cols = tax.loc[tax.group == "emo", "column"].tolist()
X = songs[emo_cols]                       # tag counts
Y = (songs[emo_cols] > 0).astype(int)     # binary multi-label targets

songs.parquet

Identifiers and metadata

Key	Type	Description
song_id	int64	Primary key, unique across all rows
alt_ids	list<int64>	Further NetEase IDs pointing at the same track. Non-empty on 8,418 rows, up to 4 entries
n_ids	int8	1 + len(alt_ids)
name	string	Track title, Chinese verbatim. 4 nulls
artist	string	5 nulls
language	dictionary<string>	42 values. instrumental 66,244 · en 42,747 · zh 39,692 · ja 13,113 · ko 2,504
duration_ms	int32, nullable	Null on 21,798 rows (12.9%) where the source carries no duration. Those rows are otherwise complete and should not be dropped
lyric	string	Present on 102,903 rows; absent almost exactly where language == "instrumental"
n_playlists	int16	Playlists containing this song. Mean 2.82, max 198

Tag columns

74 columns holding counts in the range 0–133, prefixed by group following the official NetEase tag taxonomy:

Prefix	Group	Count	Examples
lang_	Language 语种	6	lang_mandarin 华语 · lang_western 欧美 · lang_cantonese 粤语
genre_	Genre 风格	25	genre_pop 流行 · genre_rock 摇滚 · genre_new_age New Age
scene_	Scene 场景	12	scene_night 夜晚 · scene_driving 驾车 · scene_study 学习
emo_	Emotion 情感	12	emo_healing 治愈 · emo_sadness 伤感 · emo_missing 思念
theme_	Theme 主题	18	theme_acg ACG · theme_soundtrack 影视原声 · theme_ktv KTV
misc_	Outside the official taxonomy	1	misc_sexy 性感

tag_taxonomy.parquet holds the complete mapping with columns column, en, zh, group, is_emotion_label, emotion_index.

Four tags carried by the data but not listed on the official taxonomy page are grouped by meaning: 小语种 under lang, 另类/独立 and 音乐剧 under genre, 性感 under misc. 性感 sits outside emo so that the emotion label set stays at exactly 12 classes.

Emotion label vectors

Three fixed-length list columns sharing one order:

Key	Type	Description
emo_tag_counts	list<int16>[12]	Playlist tag counts
emo_ratio	list<float32>[12]	emo_tag_counts normalised to sum 1
emo_label	list<int8>[12]	1 where the count is above 0

The order lives in the Parquet schema metadata:

import pyarrow.parquet as pq, json
t = pq.read_table("songs.parquet")
EMO_ORDER = json.loads(t.schema.metadata[b"emotion_order"])
# ['Healing','Nostalgia','Excitement','Sadness','Romantic','Quiet',
#  'Happiness','Loneliness','Touching','Missing','Fresh','Relaxation']

The three list columns and the 12 scalar emo_* columns carry the same information; use whichever suits the task.

Every song carries at least one emotion label, 2.50 on average. Positive rates run from emo_healing 35.7% and emo_relaxation 32.8% down to emo_missing 11.3%.

songs_9822.parquet

A subset chosen so the 12 emotions have near-equal positive support:

	Healing	Nostalgia	Excitement	Sadness	Romantic	Quiet	Happiness	Loneliness	Touching	Missing	Fresh	Relaxation
positives	4,949	4,873	4,873	4,874	4,874	4,873	4,873	4,874	4,874	4,875	4,873	4,879

Mean 5.96 labels per song, above the 2.50 of the full table, since balancing favours multi-labelled tracks. The schema matches songs.parquet with one addition:

Key	Type	Description
unique_id	int64	Row identifier within the subset

Paired features: songs_9822_vggish.npz.

playlists.parquet

Key	Type	Description
playlist_id	int64	Join key for song_playlist.parquet
name	string	Playlist title, Chinese verbatim
tags	list<string>	Chinese tag strings, e.g. ["感动","治愈","思念"]. 75 distinct values, 1–6 per playlist, most commonly 3
n_tags	int64
confidence	dictionary<string>	normal 23,567 · vip 15,865 · official 7,181
play_count	int64
subscribed_count	int64
share_count	int64
comment_count	int64
create_time	timestamp
description	string	Curator's free text, Chinese verbatim
creator_nickname	string
creator_user_type	int64
creator_gender	int64	0 unspecified · 1 male · 2 female

tags holds Chinese strings while songs.parquet uses English column names; the zh and column fields of tag_taxonomy.parquet connect the two.

The 75 Chinese tag strings map onto 74 columns because the source spells New Age two ways, one with a regular space (U+0020) and one with a non-breaking space (U+00A0). Both denote the same tag and share the column genre_new_age.

song_playlist.parquet

Key	Type
song_id	int64
playlist_id	int64

477,216 rows. Every playlist_id resolves in playlists.parquet. 43,558 of the 46,613 playlists are referenced by at least one song.

This table makes the labels re-derivable under a different aggregation. The stored counts weight every playlist equally; the relation lets you weight by popularity or restrict to curated sources:

rel = pd.read_parquet("song_playlist.parquet")
plf = pd.read_parquet("playlists.parquet")

official = plf[plf.confidence == "official"]
sub = rel[rel.playlist_id.isin(official.playlist_id)]

weight = plf.set_index("playlist_id")["play_count"]

Aggregating the Chinese tags over each song's playlists reproduces the stored tag counts exactly.

Audio features

Both NPZ archives are keyed by song_id as a string:

import numpy as np
z = np.load("songs_169148_vggish.npz")
emb = z["1000155"]          # (31, 128) float32

File	Array per song	Description
songs_169148_vggish.npz	(31, 128) float32, range 0–255	VGGish embeddings, one 128-d frame per ~0.96 s across the middle 30 s, post-processed quantised output. Covers 169,097 songs
songs_9822_vggish.npz	(31, 128) float32	The same for the balanced subset
spectrograms_169148.npz	(64, 469) float32, range −70.8–0	Log-mel spectrogram in dB, 64 mel bands × 469 frames across the middle 30 s. Covers 169,148 songs

songs_169148_vggish.npz covers 169,097 of the 169,148 songs; check key membership before indexing.

Licence and citation

CC BY-NC-SA 4.0. Only identifiers, metadata, tags and derived features are distributed; no audio.

@article{s16k,
  doi = {10.1007/s00530-025-01701-z},
  journal = {Multimedia Systems},
  title = {A multi-label music emotion recognition dataset from NetEase Cloud Music}
}