MARBLE Benchmark Tasks and Data

Taxonomy	Task	Dataset	Prediction Head	Evaluation Metrics	Commertially Available
High-level Description	Key Detection	Giantsteps key	Multi-class	Weight score	Yes
	Music Tagging	MagnaTagATune	Multi-label	ROC-AUC & PR-AUC/AP	-
	Music Tagging	MTG Jamendo (Top50 tag)	Multi-label	ROC-AUC & PR-AUC/AP	-
	Genre classification	GTZAN	Multi-class	Accuracy	-
	Genre classification	MTG Genre	Multi-label	ROC-AUC & PR-AUC/AP	-
	Emotion Detection	emomusic	Regression	EmoV & EmoA	-
	Emotion Detection	MTG MoodTheme	Multi-label	ROC-AUC & PR-AUC/AP	-
Score-level Description	Pitch classification	NSynth	Multi-class	Accuracy	Yes
Score-level Description	Beat Tracking	GTZAN Rhythm	Seq2Seq, Binary-class	F-measure (threshold 20ms)	-
Performance-level Description	Vocal Technique Detection	VocalSet	Multi-class	Accuracy	Yes
Acoustic-level Description	Singer Identification	VocalSet	Multi-class	Accuracy	Yes
	Instrument classification	Nsynth	Multi-class	Accuracy	Yes
	Instrument classification	MTG Instrument	Multi-label	ROC-AUC & PR-AUC/AP	-
	Source Separation	MUSDB18	Seq2Seq, Regression	Source-to-distortion Ratio (SDR)	-