PaperlayerView exact revision on arXiv

Source-linked AI summary

AI-Generated Music Detection and its Challenges

Darius Afchar, Gabriel Meseguer-Brocal, Romain Hennequin

arXiv:2501.10111v1cs.SDeess.AS

TL;DR

The paper addresses detecting synthetic music as generative music becomes more accessible, while distinguishing generation artifacts from musical or codec differences remains difficult. It tests a detector using real audio and autoencoded reconstructions, reports 99.8% accuracy for amplitude spectrograms, and discusses robustness and generalisation as challenges beyond test scores.

  • Problem

    Detecting synthetic music is challenging because classifiers may learn genre or codec differences instead of generation artifacts.

    Source: A. Proposed framework · p. 2

  • Method

    The paper trains straightforward convolutional classifiers on real music and same-bitrate autoencoded reconstructions to focus on decoder artifacts.

    Source: A. Proposed framework · p. 2

  • Results

    99.8% accuracy is achieved using amplitude spectrograms, the best-performing audio representation reported.

    Source: C. First results of detection · p. 2

  • Takeaways & Limitations

    The paper presents AI-music detection as a proof of concept while urging evaluation beyond test scores, including robustness and generalisation.

    Source: §IV · p. 4

    The authors warn that straightforward detectors are not naturally robust to unanticipated cases, as attackers can evade detection and new models will appear.

    Source: C. Challenges ahead · p. 4

Abstract

from arXiv · show

In the face of a new era of generative models, the detection of artificially generated content has become a matter of utmost importance. In particular, the ability to create credible minute-long synthetic music in a few seconds on user-friendly platforms poses a real threat of fraud on streaming services and unfair competition to human artists. This paper demonstrates the possibility (and surprising ease) of training classifiers on datasets comprising real audio and artificial reconstructions, achieving a convincing accuracy of 99.8%. To our knowledge, this marks the first publication of a AI-music detector, a tool that will help in the regulation of synthetic media. Nevertheless, informed by decades of literature on forgery detection in other fields, we stress that getting a good test score is not the end of the story. We expose and discuss several facets that could be problematic with such a deployed detector: robustness to audio manipulation, generalisation to unseen models. This second part acts as a position for future research steps in the field and a caveat to a flourishing market of artificial content checkers.

I. INTRODUCTION

The paper addresses the need to identify synthetic music amid human-made content as generative music services become more accessible. It presents a general-purpose waveform-music detector and frames high detection scores as only a starting point for research on robustness and generalisation.

I. INTRODUCTION

  • The paper proposes the first general-purpose AI-generated music detector, including generated instrumental parts and focusing on waveform generators.Symbolic or MIDI-based synthesis models are left for future exploration.Source: §I · p. 1
  • Basic convolutional models achieve detection accuracy above 99%, which the paper describes as surprisingly easy to obtain.The paper presents this as a proof of concept, not as the end of evaluating a detector.Source: §I · p. 1
  • The paper urges research beyond test scores, especially on robustness to audio manipulation and generalisation to unseen generators.It connects these challenges to research questions in media-forgery detection.Source: §I · p. 1

II. AI-MUSIC DETECTION

The section introduces the detection framework, dataset, and initial detection scores as the paper’s approach to AI-music detection.

II. AI-MUSIC DETECTION

  • The section frames AI-music detection through a chosen framework, its advantages, the employed data, and initial detection scores.The passage previews these topics without specifying their details.Source: §II · p. 1
  • The framework is presented as a way to solve the AI-music detection task.The passage does not identify the framework or its specific advantages.Source: §II · p. 1
  • The section also introduces the data and first detection scores used in the study.No dataset characteristics or score values are given in this overview passage.Source: §II · p. 1

A. Proposed framework

The proposed framework aims to detect generation artifacts rather than musical content or dataset differences. It pairs real audio with autoencoded reconstructions to control musical semantics and file encoding.

A. Proposed framework

  • The framework asks whether samples reveal an artificial decoder independently of their musical content.It targets decoder artifacts such as checkerboard patterns associated with transposed convolutions.Source: A. Proposed framework · p. 2
  • Naively classifying real and synthetic music can confound generation with genre or codec differences.For example, a dataset imbalance could teach a classifier to distinguish classical music from rap and pop instead of detecting forgery.Source: A. Proposed framework · p. 2
  • The method autoencodes real tracks and stores reconstructions at the originals’ bitrate, controlling for musical semantics and file encoding.The classifier is then intended to distinguish real audio from its reconstructed counterpart by generation artifacts.Source: A. Proposed framework · p. 2

B. Considered dataset and music generators

The study uses the FMA medium split and a selected set of neural-codec and other decoder reconstructions. The dataset pairs real tracks with reconstructions while preserving musical content and bitrate.

B. Considered dataset and music generators

  • The FMA medium split provides 25,000 tracks across 16 genres, encoded as MP3 at varied bitrates and processed at 44.1 kHz.The dataset is selected for openness and reproducibility; the study limits its scope to the medium split for size reasons.Source: B. Considered dataset and music… · p. 2
  • The study considers Encodec, DAC, Musika, and GriffinMel, with several bitrate or melband configurations.The availability of trained models constrained decoder selection; some music checkpoints and closed-source systems were unavailable.Source: B. Considered dataset and music… · p. 2
  • The resulting dataset contains 250,000 tracks split 70/10/20 across train, validation, and test.Each reconstruction retains the original track’s semantic content and bitrate while adding decoder-specific artifacts.Source: B. Considered dataset and music… · p. 2

C. First results of detection

Straightforward convolutional models perform strongly, with input representation affecting results more than architecture. Amplitude spectrograms give the best reported accuracy, while phase-only input also performs well.

C. First results of detection

  • The straightforward convolutional setup initially achieves test accuracies over 90%, prompting sanity checks rather than more complex models.The model alternates convolution and pooling layers.Source: C. First results of detection · p. 2
  • 99.8% accuracy is the best reported result, achieved with amplitude spectrograms.The authors report that architectural choices affect performance less than input preprocessing.Source: C. First results of detection · p. 2
  • A phase-only model also yields high scores, despite phase often being considered less efficient than amplitude representation.The paper reports consistent per-class performance across autoencoders.Source: C. First results of detection · p. 2

D. Generalisation to full music generators

The authors test whether detection extends beyond autoencoded training examples using text-prompted MusicGen tracks. They report a 99.9% detection score on snippets from those tracks.

D. Generalisation to full music generators

  • The MusicGen experiment tests detection on text-prompted synthetic music unseen during training, rather than autoencoder reconstructions.The authors frame it as a sanity check of their concern that the classifier might detect autoencoder fingerprints.Source: D. Generalisation to full music… · pp. 2–3
  • 99.9% detection score is achieved on 2,500 snippets from 50 MusicGen tracks, totaling 25 minutes.The test samples are generated from text prompts and were not autoencoded or related to FMA.Source: D. Generalisation to full music… · pp. 2–3
  • The authors prioritize examining the implications of convincing scores over further architecture search.They say the results could make it feel as if the task were solved, but question whether that confidence is warranted.Source: D. Generalisation to full music… · p. 3

III. CAVEATS ON AI-MUSIC DETECTORS

The paper cautions that strong detector scores do not settle whether AI-music detection is realistic or reliable. It highlights robustness, generalization, and the trustworthiness of performance claims as concerns.

III. CAVEATS ON AI-MUSIC DETECTORS

  • Commercial AI-music detectors are closed-source and lack associated research publications, preventing rigorous study of their claims.The paper connects this concern to closed-source AI-text detectors and difficulties verifying their performance.Source: §III · p. 3
  • The discussion identifies robustness to audio manipulation and generalization to unknown encoders as challenges for realistic, reliable detection.It also raises the trustworthiness of performance numbers and mentions open source and interpretability as validation considerations.Source: §III · p. 3

A. Robustness to manipulations

The authors examine robustness to common audio manipulations and find that some substantially reduce detection performance. They suggest this may reflect reliance on autoencoder-specific artifacts, raising concern about generalization to unseen generators.

A. Robustness to manipulations

  • The authors describe data shifts such as reencoding, adding music to video, time stretching, or pitch shifting as ways outputs may differ from a generator’s immediate output.They note that some users could apply transformations strategically to evade detection.Source: A. Robustness to manipulations · p. 3
  • The study tests common transformations including pitch shifts of ±2 semitones, time stretch of 80–120%, EQ, reverb, white noise, and MP3, AAC, and Opus reencoding at 64 kbps.Attacks from more advanced users, such as adversarial attacks, are left for future work.Source: A. Robustness to manipulations · p. 3
  • Detection performance drops drastically under pitch shifts, white noise, and codec reencoding, while remaining robust to time stretch, EQ, and reverberation.The authors report using an amplitude-based model on transformations unseen during training; they say the reason for robustness to some manipulations is unclear.Source: A. Robustness to manipulations · p. 3
  • The authors suggest the model detects autoencoder-specific artifacts and defaults to the real class when it finds none, or when manipulation makes them unrecognizable.They report that some scores drop to almost zero while real-class scores are unaffected.Source: A. Robustness to manipulations · p. 3
  • Because real music may be more diverse and complex than autoencoded generations, the authors warn that the detector may not generalize to unseen generators.They connect this concern to machine-learning models’ bias toward simple solutions.Source: A. Robustness to manipulations · p. 4

B. Encoder generalisation

Encoder generalisation is strong within decoder families but almost always fails across families, leaving transfer to unseen models a central weakness.

B. Encoder generalisation

  • Inter-family transfer is almost always near zero, including GriffinMel→DAC, and the real class may act as the default.The authors connect this pattern to the detector’s poor robustness to unseen manipulations.Source: B. Encoder generalisation · p. 4
  • Encodec 24kbps reconstruction transfers well to 6kbps and 3kbps within the same family.The authors conclude that including every parameterisation of an autoencoder may not be necessary for within-family detection.Source: B. Encoder generalisation · p. 4
  • Training on a higher bitrate seems to transfer better to lower bitrates, though the authors caution that this is not straightforward to assert.They suggest the considered models’ RVQ formulation could contribute.Source: B. Encoder generalisation · p. 4

C. Challenges ahead

The authors highlight limits in robustness, calibration, and detector-centered regulation. They call for evaluation beyond optimized settings and attention to broader responses to AI-music risks.

C. Challenges ahead

  • Fine-tuning on audio manipulations or a new decoder could restore high accuracy, but the authors argue that unseen manipulations and generation methods will remain.They did not train on the manipulations used to study natural robustness.Source: C. Challenges ahead · p. 4
  • Straightforward detectors are not naturally robust to unanticipated cases, motivating partial-knowledge evaluation and regular patching.The authors describe detection as a continuing cat-and-mouse game with new evasion methods and models.Source: C. Challenges ahead · p. 4
  • The model’s output probabilities should not be read as percentages of AI content, and interpretability and calibration merit further study.The authors point to mixed real and synthetic stems as a calibration concern and interpretability as useful for debugging false positives.Source: C. Challenges ahead · p. 4
  • Detector-based regulation may overlook other parts of the AI supply chain, while watermarking is an imperfect alternative.The authors suggest regulating large technology actors may be more efficient than responding to generations after they spread.Source: C. Challenges ahead · p. 4

IV. CONCLUSION

The paper presents AI-generated music detection as a first study and proof of concept, while arguing that accuracy alone is insufficient. It identifies robustness and generalisation as challenges and outlines further research directions.

IV. CONCLUSION

  • The paper presents itself as the first study of AI-generated music detection and reports that forged content is surprisingly easy to detect.The authors caution that a good accuracy score is not the end of the story.Source: §IV · p. 4
  • The authors recommend considering robustness to manipulation and generalisation to unseen settings alongside detection accuracy.Source: §IV · p. 4
  • Future work includes fine-tuning or updating detectors for new generators, further augmentation, adversarial defense, interpretability, and more realistic stem mixing and audio engineering.Source: §IV · p. 4
Loading 2501.10111v1…