Tracing Inputs, Verifying Outputs: Validating Attribution in Music Generation

Taejun Kim; Wonil Kim; Jongmin Jung; Hyeongseok Wi; Sangeun Kum; Keunhyoung Luke Kim; Taehyoung Kim; Dongjoo Moon

cs.SDOct 7, 2026

Tracing Inputs, Verifying Outputs: Validating Attribution in Music Generation

Authors: Taejun Kim, Wonil Kim, Jongmin Jung, Hyeongseok Wi, Sangeun Kum, Keunhyoung Luke Kim, Taehyoung Kim, Dongjoo Moon, +5 more

Organizations: Neutune, Seoul, South Korea · KAIST, Daejeon, South Korea

Abstract

How can we verify whose music contributed to an AI-generated output? This paper demonstrates how input-based attribution can provide verifiable evidence of which audio sources were used in a generation and whether they shaped the output. To do so, we condition the generation solely on audio without any text input, then trace the inputs behind each output, and establish their musical effect. In prompt adherence tests and controlled input swaps, the stems generated by our generator, MixAudio, follow the prompt audio in timbre and the context audio in harmony. Yet these outputs may still reproduce training data not supplied as inputs. We therefore audit memorization with our musical version identification model, musicDNA, and find few reproductions outside the input records. On human-judged cases within the flagged pool, it achieves higher precision and recall than the other tested memorization detectors. The two evaluations suggest that input records and output analysis provide complementary evidence for attribution, on which rights-holder reporting and compensation can draw as the AI music economy takes shape. Audio examples are available at https://neutune.github.io/attr2027demo/

Figures & tables

Figure 1 : Attribution by conditioning provenance. A single pass generates any subset of a track’s stems, conditioned exclusively on registered audio, with no text input: one prompt recording per stem and an optional shared context. The two rights panels on the left are the rights involved at generation. The dashed box is added when the context is mixed into the released master.

		MERT (Style)			MFCC (Timbre)
Source	System	Matched	Random	$\Delta$	Matched	Random	$\Delta$
Training	MixAudio	$0.550$	$0.014$	$\mathbf{0.535}$	$0.914$	$0.554$	$\mathbf{0.360}$
Training	ACE	$0.549$	$0.121$	$0.428$	$0.838$	$0.602$	$0.236$
Validation	MixAudio	$0.516$	$0.000$	$\mathbf{0.516}$	$0.911$	$0.550$	$\mathbf{0.361}$
Validation	ACE	$0.534$	$0.123$	$0.411$	$0.831$	$0.602$	$0.229$
MoisesDB	MixAudio	$0.582$	$0.124$	$\mathbf{0.457}$	$0.920$	$0.599$	$\mathbf{0.321}$

Table 1: Prompt adherence. Similarity between generated stems and prompts. Matched : each stem against the prompt it was generated from. Random : the same generations scored against another track’s prompt.

\Delta

= Matched

-

Random. Bold: larger

\Delta

per source.

(a) Prompt replaced
		MixAudio		ACE
Hypothesis	Metric	Base	Swap	Base	Swap
Timbre moves to the substituted prompt	MFCC	$0.63$	$0.89$	$0.69$	$0.85$
Timbre leaves the original prompt	MFCC	$0.92$	$0.64$	$0.88$	$0.70$
Harmony stays with the original context	Chroma	$0.20$	$0.22$	$0.10$	$0.06$
(b) Context replaced

Table 2: Conditioning swaps. Each value is the similarity between a generated stem and the prompt or context marked in its row. Base : the base generation. Swap : the regeneration after the swap; the other conditioning input is held fixed. Averaged over the four dataset sources.

Figure 2 : How listeners judge a flagged stem, read top to bottom. Orange: Type I and II, the material unnamed memorizations; blue: named (A) or minor unnamed (D) memorization. Badges:

n

\cdot

% of the band

\cdot

% of the 994 queries, shown for the musicDNA band over both splits (59 stems).

	Validation (unseen)				Training (registered)
Case	musicDNA	CLEWS	CLAP	MERT	musicDNA	CLEWS	CLAP	MERT
Flagged	25	25	25	25	34	35	31	24
Type I $\cdot$ Recording	0	0	0	0	1 (3%)	1 (3%)	0	0
Type II $\cdot$ Composition	0	0	0	0	0	0	0	0
A $\cdot$ Duplicate	9 (36%)	7 (28%)	1 (4%)	3 (12%)	19 (56%)	14 (40%)	0	7 (29%)
B $\cdot$ Unrelated	16 (64%)	12 (48%)	24 (96%)	22 (88%)	13 (38%)	16 (46%)	31 (100%)	17 (71%)

Table 3: Listener judgments on the flagged stems, per retriever and split, in the categories of Figure 2 . Percentages are shares of each column’s flagged stems.

Figure 3 : Detection performance of memorization metrics. One dot per human-judged stem, scored by all six detectors. Rows follow the categories of Figure 2 , in the same colors: colored dots are confirmed memorizations (Type I, A), grey dots are not (B, C). More similar is to the right, and the dashed line is each detector’s detection threshold. The Null row (grey band) marks chance-level similarity: the 5–95% range of scores for unseen validation stems. Vertical jitter within a row carries no meaning.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Source	Tracks	Sections	Instances	Target stems	Contextless	Hours
Train	1922	2073	2073	5030	515	7.1
Val	168	2086	2086	4960	518	6.9
MoisesDB	112	1058	1058	2266	282	5.6
MUSDB18	150	1593	1593	2778	458	8.3

Appendix

Table 4: Evaluation set composition. One instance per section; target stems are the stems each system generates; contextless instances have no context mix. Hours: total section audio.

System	Prompt	Context	Stems	Why not a baseline
ACE-Step 1.5, Lego ( Gong et al., 2026 )	audio, per stem	✓	✓	baseline
MTMusicLDM ( Karchkhadze et al., 2026 )	text	✓	✓	no audio prompt
MSDM ( Mariani et al., 2024 )	–	inpainting	4 fixed	no audio prompt
YuE ( Yuan et al., 2026 )	audio	–	2 fixed	no context
MusicGen-Melody ( Copet et al., 2023 )	chroma	–	–	no context
Stable Audio Open ( Evans et al., 2025 )	audio-to-audio	–	–	no context

Appendix

Table 5 : Candidate baselines. Research systems with released weights (top) and commercial services (bottom). What each system takes as the prompt for the generated part and whether it takes existing audio as a context and produces stems (✓ yes, – no).

(a) Prompt time-scaled ( $\times$ 0.9 / 1.1)
Hypothesis	Metric	Oracle	MixAudio	ACE
Output tempo unchanged	Tempo accuracy	$100\%$	$95\%$	$91\%$
Harmony stays with the context	Chroma $\Delta$	$\approx 0$	$-0.003$	$+0.003$
Timbre unchanged	MFCC $\Delta$	$\approx 0$	$-0.005$	$-0.005$
(b) Prompt pitch-shifted ( $\pm 2$ st)
Harmony stays with the context	Chroma $\Delta$	$\approx 0$	$-0.005$	$-0.025$

Appendix

Table 6: Tempo and pitch transformations. One conditioning input is time-scaled or pitch-shifted by a known amount and the stem regenerated, while the other conditioning input is held fixed. Oracle is the value perfect following would yield.

System	Source	FAD ( $\times 10^{-3}$ ) $\downarrow$	Human ratings (MOS 1–5) $\uparrow$
			Quality	Prompt	Context
MixAudio	Validation	$0.596$	$2.81\pm 0.29$	$3.36\pm 0.32$	$3.45\pm 0.39$
	MoisesDB	$0.997$	$2.77\pm 0.33$	$3.34\pm 0.33$	$3.43\pm 0.42$
	MUSDB18	$1.010$	$2.80\pm 0.27$	$3.18\pm 0.31$	$3.20\pm 0.30$
	All	$0.868$	$2.80\pm 0.17$	$3.29\pm 0.19$	$3.36\pm 0.22$
ACE	Validation	$0.935$	$2.36\pm 0.25$	$2.73\pm 0.34$	$2.56\pm 0.44$

Appendix

Table 8: FAD and listener ratings of generated stems, and listener ratings of the real stems of the same cases. FAD: lower is better. Ratings: mean opinion scores from the listening test (1–5) with 95% confidence intervals, higher is better. Quality, Prompt, and Context are sound quality, match to the prompt audio, and fit with the context.

Figure 4 : Listener ratings. Mean opinion scores over all 48 cases, the All rows of Table 8 , with their 95% confidence intervals.

Explore similar work

ARIA: A Diagnostic Framework for Music Training Data Attribution
May 15, 2026Changheon Han, Ashkan Panahi, Kıvanç TatarMusic Structure Analysis Training Data
Department of Computer Science and Engineering, Chalmers University of Technology and University of Gothenburg Gothenburg, Sweden
Towards AI-Generated Music Plagiarism Detection as a Version Identification Problem
Oct 6, 2026Fotis Koutsikos, Ioannis Prokopiou, Spyridon Kantarelis +6Copyright Defenses Generative Artificial Intelligence
National Technical University of Athens, Athens, Greece · Athens University of Economics and Business, Athens, Greece · Orfium, Athens, Greece +1
ArtifactNet: Detecting AI-Generated Music via Forensic Residual Physics
Apr 17, 2026Heewon OhGenerative Music Ai-Generated Content Detection
Intrect / MARTE Lab, Dongguk University, Seoul, South Korea

May 15, 2026, cs.SDJ/K move · Enter open · S save

ARIA: A Diagnostic Framework for Music Training Data Attribution

Changheon Han, Ashkan Panahi, Kıvanç Tatar

Department of Computer Science and Engineering, Chalmers University of Technology and University of Gothenburg Gothenburg, Sweden

Music Structure Analysis Training Data Copyright Defenses Diagnostic Framework

Open paper AskPDF arXiv

Training data attribution (TDA) for music generation must answer two questions that copyright analysis requires, namely which training songs influence a generated output and along which musical aspects the influence operates. Existing methods reduce influence to a single scalar, without revealing which musical aspects are dominant in that influence. We propose ARIA, a framework that decomposes attribution along musical aspects (five for symbolic music, three for audio) and pairs the decomposition with reliability diagnostics computed from the segment-level score matrix. It measures within-group similarity among the top-K attributed tracks against random reference groups drawn from the training pool, and diagnoses the score matrix through its singular value decomposition and column statistics. On a symbolic-music model where attribution ground truth is available through counterfactual retraining, the reliability diagnostics rank four attribution methods identically to that ground truth. On an audio music generation model, ARIA reveals attribution behaviors that vary substantially across TDA methods, flags score matrices whose retrieved tracks are nearly identical across queries rather than reflecting per-query attribution, and characterizes embedding-similarity retrieval baselines by the musical aspect each encoder surfaces. Together, ARIA produces per-aspect attribution evidence aligned with the musical aspects considered under the idea-expression distinction in copyright analysis.

Abstract

Explore similar work

ARIA: A Diagnostic Framework for Music Training Data Attribution

Towards AI-Generated Music Plagiarism Detection as a Version Identification Problem

ArtifactNet: Detecting AI-Generated Music via Forensic Residual Physics