DaveLoay/bamm-cnn-detectors

BAMM CNN detectors: CNN Clean and CNN Broadcast

Two binary classifiers that estimate whether a piece of music is AI-generated (Suno v3.5) or human-made. They are the baselines of the ISMIR 2026 paper Assessing AI-generated music detection in real-world broadcast monitoring (arXiv:2608.07359), which introduces the BAMM dataset of real TV broadcast clips.

Code (training, evaluation, inference): https://github.com/DaveLoay/BAMM-dataset

Folder	Model	Trained on
cnn_clean/	CNN Clean	Clean foreground music: FMA-medium (human) vs SONICS Suno v3.5 (AI)
cnn_broadcast/	CNN Broadcast	The same music, 70% of each class mixed with LibriSpeech speech at a random SNR in [−30, +30] dB, 30% left clean, all re-encoded as 8 kHz / 40 kbps MP3

Each folder holds:

File	Content
model.h5	Keras model, legacy HDF5 format, no optimizer state (6.6 MB, 1.64 M parameters)
config_snapshot.yaml	The full training configuration, including the feature parameters inference must use
training_history.json	Per-epoch loss, accuracy and ROC-AUC
model_summary.txt	Layer-by-layer summary

Usage

With the GitHub repository:

git clone https://github.com/DaveLoay/BAMM-dataset.git && cd BAMM-dataset
pip install -e .
python scripts/download_weights.py
python -m pipelines.predict --model-path weights/cnn_broadcast/model.h5 --input clip.mp4

Or with the single-file inference.py from this repository:

pip install "tensorflow>=2.11,<2.16" "librosa>=0.10" "numpy<2" pyyaml huggingface_hub
python inference.py --model cnn_broadcast song.mp3 clip.mp4

Both need ffmpeg on the PATH. Any format FFmpeg can decode works, including MP4 video.

The output is P(AI), the mean of the model's scores over non-overlapping 5 s windows. The paper classifies a clip as AI-generated when P(AI) >= 0.5.

Input and preprocessing

The models expect exactly this pipeline (implemented in both options above):

1. FFmpeg transcode to mono MP3, 8 kHz, 40 kbps
2. Power mel spectrogram: STFT 2048, hop 1024, window 2048, 128 bands from 20 Hz to 4 kHz
3. librosa.power_to_db(S, ref=np.max), then (S + 4.0) / 3.0
4. Non-overlapping 5 s windows of 38 frames, input shape (38, 128, 1); clips shorter than 5 s are zero-padded

Architecture

Six blocks of 3×3 Conv2D (16, 32, 64, 128, 256, 512 filters, ReLU), each followed by 2×2 max pooling and batch normalisation; global average pooling; dropout 0.2; dense 128 (swish); dropout 0.2; one sigmoid unit. The design follows Afchar et al., AI-generated music detection and its challenges.

Training

Adam (learning rate 1e-3, halved on plateau), batch size 50, binary cross-entropy, up to 100 epochs with early stopping on validation loss (patience 15). Each epoch draws random 5 s crops from every track: 5 per track for CNN Clean, 3 for CNN Broadcast. The stratified 70/20/10 split uses seed 123. The saved checkpoint is the one with the best validation ROC-AUC.

Evaluation

Results from scripts/reproduce_paper.py in the GitHub repository (F1 at threshold 0.5, ROC-AUC):

Scenario	Unit	n	CNN Clean F1	CNN Clean ROC	CNN Broadcast F1	CNN Broadcast ROC
CFM	clip	952	0.992	0.998	0.992	0.996
STB	clip	3,294	0.342	0.910	0.661	0.926
RTB	clip	5,498	0.186	0.761	0.472	0.796
RTB	DMD segment, all	8,653	0.172	0.707	0.422	0.775
RTB	DMD segment, foreground	3,359	0.317	0.782	0.519	0.858
RTB	DMD segment, background	5,225	0.039	0.667	0.338	0.745

• CFM, Clean Foreground Music: the 476 human / 476 AI clean reference tracks AI-OpenBMAT was built from (not publicly available)
• STB, Synthetic TV Broadcast: the 3,294 AI-OpenBMAT mixtures
• RTB, Real TV Broadcast: BAMM, by clip and by Deep Music Detector segment

The paper's Table 1 uses the clip-level F1 and the all-segments ROC-AUC for RTB. The foreground and background rows match its Figure 3.

Limitations

• The AI class is Suno v3.5 only. Other generators and later Suno versions were not part of training.
• On real broadcast audio both models degrade a lot, mostly by missing AI music (false negatives), and more so when the music is in the background. The paper's conclusion is that these detectors are not reliable enough for broadcast monitoring on their own.
• The models only see content below 4 kHz, at 8 kHz / 40 kbps.
• Do not use a single score to make decisions about a person or a work. Treat the output as one signal among others.

License

The weights are released under CC BY-NC 4.0 (attribution, non-commercial). They were trained on FMA-medium, SONICS and LibriSpeech; check those datasets' terms for your use. The code in the GitHub repository is MIT-licensed.

Citation

@inproceedings{lopezayala2026bamm,
  title     = {Assessing {AI}-generated music detection in real-world broadcast monitoring},
  author    = {L{\'o}pez-Ayala, David and Garcia de la Cruz, Fernando and Zinemanas, Pablo
               and Molina, Emilio and Rocamora, Mart{\'i}n},
  booktitle = {Proc. of the 27th Int. Society for Music Information Retrieval Conf. (ISMIR)},
  address   = {Abu Dhabi, UAE},
  year      = {2026}
}