Assessing AI-generated music detection in real-world broadcast monitoring

The paper is authored by David López-Ayala, Fernando Garcia de la Cruz, Pablo Zinemanas, Emilio Molina, and Martín Rocamora, affiliated with the Music Technology Group at Universitat Pompeu Fabra (Barcelona, Spain) and BMAT Licensing S.L. (Barcelona, Spain). It was published in the Proceedings of the 27th International Society for Music Information Retrieval Conference (ISMIR), Abu Dhabi, UAE, 2026, and is available on arXiv (arXiv:2608.07359v1).

The paper addresses the problem of detecting AI-generated music in real-world broadcast monitoring. The abstract states: "The proliferation of AI-generated music in broadcast media raises concerns about transparency and fair compensation, but reliable detection under real broadcast conditions remains unresolved. Existing studies report substantial performance degradation in this domain, yet their evaluations are limited to synthetic broadcast data."

To address this gap, the authors introduce BAMM (Broadcast AI-Music Monitoring), a 40-hour dataset of real-world television recordings containing both AI-generated and human-made music. The paper compares clean-trained and broadcast-trained CNN variants across three progressively more challenging scenarios: Clean Foreground Music (CFM), Synthetic TV Broadcast (STB), and Real TV Broadcast (RTB). Key findings include near-perfect performance on CFM for both models, substantial degradation under synthetic broadcast conditions, and further degradation on RTB with substantial score overlap between AI-generated and human-made music. The authors conclude that these results expose a critical domain gap and show that current training approaches on CNN-based detectors remain insufficient for reliable AI-generated music detection in broadcast monitoring.

The introduction highlights the rapid lowering of barriers to music production through generative models, alongside technical and legal challenges concerning training data provenance, training data replication, the definition of authorship, and compensation frameworks. These have resulted in lawsuits against AI-music companies such as Suno and Udio, licensing deals between generative-music platforms and rightsholders, and policies such as the EU AI Act introducing transparency requirements.

The urgency is underscored by Deezer's report that, for the first time, 50% of its daily uploads were detected as fully AI-generated, and similar trends observed in broadcast media, with companies such as BMAT reporting a growing presence of AI-generated music in television recordings.

While existing detectors achieve near-perfect accuracy on isolated, high-fidelity tracks, they are often not robust to domain shifts and common audio transformations, such as random pitch shifting, time stretching, EQ, reverb, among others. The broadcast environment introduces significant complexities: music segments are typically short, frequently backgrounded by dominant speech or sound effects, and subjected to transmission constraints. Recent benchmarks have shown that even in synthetic mixtures, these factors substantially degrade performance on state-of-the-art architectures, yet the extent to which AI detectors work in real-world broadcast conditions remains under-explored.

BAMM is described as a novel dataset comprising 40 hours of real-world broadcast content curated for AI-generated music detection. It contains "approximately 20 hours of AI-generated music and 20 hours of human-made music, stored as.mp4 clips. Each sample captures original broadcast audio and video, preserving the authentic acoustic degradations and production contexts (e.g., music in the background, sound effects) found in practice."

The clips were sampled from a global broadcast archive that monitors over 4,200 television channels worldwide, with all samples aired between January 2025 and March 2026 and durations ranging from 5 s to 60 s. Reflecting industrial monitoring constraints, the audio is monophonic, sampled at 8 kHz, and encoded with AAC-LC at bitrates of at least 40 kbps, which pose a significant worst-case challenge for detection models.

The dataset was built via an automated pipeline combining reference track curation, audio fingerprinting, and multi-stage filtering.

Reference Track Selection: The pipeline uses a candidate pool of clean MP3 reference tracks categorized by an ensemble of five AI-music detectors to ensure trustworthy labels. For the human-made class, reference tracks were restricted to releases between January 2020 and December 2022, ensuring that they predated Suno v3.5 and other publicly available or commercial generative music systems capable of producing music of comparable quality. For the AI-generated class, a candidate track was included only if it exceeded the calibrated decision thresholds across all five detectors, requiring unanimous agreement from the ensemble.

The detector ensemble comprises: (1) a publicly available CNN (referred to as CNN Clean in the paper, based on Afchar et al.'s architecture, trained on publicly available data, operating on 5-second windows as mel-spectrograms from audio resampled to 8 kHz); (2) the publicly available SpecTTTra alpha-5s detector trained on the SONICS Suno v3.5 subset, operating on spatio-temporal tokens at 16 kHz; (3-5) three internal models including a CNN based on Afchar et al.'s architecture operating at 16 kHz, a logistic regression classifier on artifact-fingerprint features, and an MLP on the same features.

All five models were trained as binary classifiers under clean foreground music conditions to distinguish human-made from AI-generated tracks (Suno v3.5), and all achieved F1 scores above 98%. Conservative calibration achieved a zero false positive rate using 5,000 tracks from Da-TACOS and a control set of 500 AI-generated tracks from the clean subset of AI-OpenBMAT. Temporal validation of the selection strategy (Figure 2) shows detection scores consistently low before the Summer 2024 release date of Suno v3.5, increase markedly after its introduction, and later decrease over time, consistent with real-world adoption of Suno v3.5 and newer models.

Fingerprint and Broadcast Retrieval: The authors used a private implementation of audio fingerprinting based on landmarks to locate matches in our TV broadcast archive, searching recordings aired between January 2025 and March 2026 and extracting matched segments directly from the original TV emissions (archived at 8 kHz and 40 kbps).

DMD Filtering: A Deep Music Detector (DMD), based on the approach proposed by Meléndez et al., classified audio segments as foreground music, background music, or speech. Only clips labeled as containing music were kept, thereby removing sound effects and incorrect fingerprint matches.

The benchmark examines how the training domain affects CNN-based detectors designed to identify AI-generated music in real broadcast settings, comparing two variants of the Afchar et al. architecture both operating on 5-second windows but trained under different conditions: clean foreground music and broadcast-oriented data. The focus is on Suno v3.5, as model-agnostic detection remains an open problem.

The CNN has six convolutional layers with filter sizes [16, 32, 64, 128, 256, 512] and kernel size 3, followed by average pooling and two fully connected layers. Audio is converted to mono, resampled to 8 kHz, transformed via STFT (window 2048, hop 1024) into 128 Mel bands spanning 20 Hz to 4 kHz, converted to log dB, and standardized with a fixed global mean of −4.0 and standard deviation of 3.0.

Trained on clean foreground music only. The human-made class is drawn from the FMA-medium dataset, comprising approximately 25k songs, while the AI-generated class is built from the Suno v3.5 subset of SONICS, comprising approximately 19k songs. For each batch, five 5 s snippets are randomly sampled from each track.

Same architecture and input representation, but trained on data designed to emulate broadcast conditions by mixing music and speech. Music sources are the same as CNN Clean; speech comes from the 360-hour clean training partition of LibriSpeech. The training set uses a controlled ratio of 70% mixed samples and 30% clean samples for each class, with speech mixed with the music signal at a randomly sampled SNR between −30 dB and +30 dB. Files are exported in mono at 8 kHz and encoded as.mp3 at 40 kbps before entering the shared preprocessing pipeline.

1. Clean Foreground Music (CFM): Uses clean tracks from AI-OpenBMAT, containing 476 human–AI pairs with a sampling rate of 22.05 kHz and a bitrate of 353 kbps, where AI tracks were generated from human tracks using Suno's extend function, making pairs closely matched in musical characteristics such as style, timbre, pitch, and key.

2. Synthetic TV Broadcast (STB): Uses the full AI-OpenBMAT dataset, 3,294 tracks, totaling 54.9 hours of synthetic broadcast material, with a sampling rate of 22.05 kHz and a bitrate of 353 kbps, created from the same clean references using controlled mixtures designed to accurately emulate broadcast conditions.

3. Real TV Broadcast (RTB): Uses the full BAMM dataset, comprising 40 hours of real TV broadcast recordings, with approximately 20 hours per class, at a sampling rate of 8 kHz and a bitrate of 40 kbps using AAC-LC encoding. This extends previous broadcast-oriented evaluations by shifting from synthetic approximations to real television content.

The results are presented in Table 1:

| Scenario | CNN Clean F1 | CNN Clean ROC | CNN Broadcast F1 | CNN Broadcast ROC |

|---|---|---|---|---|

| CFM | 0.992 | 0.998 | 0.992 | 0.996 |

| STB | 0.342 | 0.909 | 0.661 | 0.926 |

| RTB | 0.186 | 0.707 | 0.472 | 0.775 |

Both models "achieve F1-scores above 99%, demonstrating that the artifacts learned from low-quality training data are still present in the high-quality data typically used for this task and remain detectable even when considering only the lower-frequency components of the audio spectrum."

The transition to mixed-audio conditions produces a significant reduction in detection performance for both models. The clean-trained model suffers the greatest degradation, demonstrating that artifacts learned from isolated music do not fully transfer when partially masked by the surrounding broadcast context. CNN Broadcast alleviates this degradation by learning from similar mixed conditions, but its limited performance suggests that the problem is not only related to the training distribution but also to an intrinsic reduction in artifact visibility arising from the broadcast scenario's mixture process.

RTB is the most challenging setting in this work, with both models performing worse than in STB, indicating that the synthetic setting does not fully capture real broadcast conditions. CNN Clean shows the largest degradation; in addition to reduced audio quality (8 kHz, 40 kbps AAC-LC), sound effects, transitions, and diverse mixing strategies can further obscure AI-related artifacts.

Figure 3 (ROC curves) shows both models retain some discriminative ability and perform better on foreground than on background music, with CNN Broadcast achieving AUC 0.858 on foreground vs. 0.745 on background (all clips: 0.775), and CNN Clean achieving 0.782 on foreground vs. 0.667 on background (all clips: 0.707). This confirms masking is a major source of degradation, but the limited performance on foreground segments also indicates that degraded audio quality and other broadcast sounds affect artifact visibility.

Figure 4 (score distributions) reveals "the main difficulty is detecting AI-generated music rather than rejecting human content. While human samples are generally assigned scores near zero, many AI samples also receive low scores, particularly in background music. Overall, broadcast-simulated training improves robustness, but it does not fully overcome the effects of real broadcast conditions."

The paper concludes that "CNN-based detectors trained on clean foreground music or simulated broadcast data retain some discriminative capacity under controlled conditions, but their performance degrades substantially as the evaluation scenario becomes more realistic. Although performance remains limited for both foreground and background music, the models perform slightly better when the music is in the foreground, suggesting that current detectors are not reliably capturing AI-generated music artifacts, and that the task becomes even more difficult when the musical signal is less salient or embedded within complex audio mixtures."

The authors highlight the need for more robust detection methods capable of modeling the acoustic and contextual complexity of real broadcast media, and position BAMM as "a benchmark for this direction by exposing failure modes not fully captured by clean-music datasets and encouraging the development of models that can detect AI-generated music in challenging, real-world broadcast environments."

The work was supported by the Cátedra IA y Música project (TSI-100929-2023-1), funded by the Secretaría de Estado de Digitalización e Inteligencia Artificial, the European Union-Next Generation EU funds, BMAT Music Innovators, and the IMPA project (PID2023-152250OB-I00) funded by MCIU/AEI/10.13039/501100011033/FEDER, UE. The dataset (.mp4 recordings) is publicly available on Zenodo, with baseline code and evaluation scripts on GitHub (https://github.com/DaveLoay/BAMM-dataset.git).

1. Robust broadcast-domain detection: The system can detect AI-generated music in real TV broadcasts at 8 kHz mono / 40 kbps AAC-LC, maintaining high F1 on foreground and background music rather than collapsing to near-chance levels (current best: 0.472 F1 / 0.775 ROC on RTB).

2. Acoustic-masking-aware architecture: The system explicitly models speech/music mixtures and sound effects, using source separation or masking-invariant representations so detection works even when music is backgrounded (target: >0.85 ROC on background music, vs. current 0.745).

3. Domain-adaptive training: The system is trained with a mix of clean, synthetic-broadcast, and real-broadcast data, using domain-adversarial or self-supervised pretraining on unlabeled TV audio to close the synthetic-to-real gap (target: reduce RTB degradation vs. STB from 0.15 ROC drop to <0.05).

4. Low-bitrate artifact preservation: The system uses front-end enhancements (e.g., bandwidth extension, denoising, or spectral restoration) so that AI artifacts survive heavy compression; it detects generative traces in the 20 Hz–4 kHz band without relying on high-frequency content.

5. Temporal context integration: Instead of isolated 5-second windows, the system aggregates predictions over 10–60 s clips using attention or HMM smoothing, leveraging production context (transitions, jingles, foreground/background changes) to disambiguate low-score AI samples from human samples.

6. Reliable confidence/uncertainty estimation: The system outputs calibrated probabilities with uncertainty, flagging low-confidence AI clips for human review rather than forcing binary decisions—addressing the substantial score overlap observed between AI and human music on real broadcasts.

7. BAMM-aligned evaluation: The system is benchmarked on real TV recordings (not just synthetic mixtures), with separate reporting for foreground, background, and full-clip conditions, ensuring progress translates to operational broadcast monitoring rather than clean-set accuracy.

The proliferation of AI-generated music in broadcast media raises concerns about transparency and fair compensation, but reliable detection under real broadcast conditions remains unresolved. Existing studies report substantial performance degradation in this domain, yet their evaluations are limited to synthetic broadcast data. To address this gap, we introduce BAMM (Broadcast AI-Music Monitoring), a 40-hour dataset of real-world television recordings containing AI-generated and human-made music. We compare clean-trained and broadcast-trained CNN variants across three progressively more challenging scenarios: Clean Foreground Music (CFM), Synthetic TV Broadcast (STB), and Real TV Broadcast (RTB). Both models achieve near-perfect performance on CFM but degrade substantially under synthetic broadcast conditions. Broadcast-oriented training improves robustness compared with clean training, although performance remains limited. On RTB, evaluated using BAMM, both models degrade further and show substantial score overlap between AI-generated and human-made music. These results expose a critical domain gap and show that current training approaches on CNN-based detectors remain insufficient for reliable AI-generated music detection in broadcast monitoring.