orphea-detect-v1

Orphea Detect v1

Binary classifier: AI-generated music vs human-made music. Built for independent artists. Open source, MIT licensed.

Model

MobileNetV3-Small fine-tuned on mel-spectrograms (128 bands, first 30 seconds of audio).

Training data

• AI: MusicGen-small generated clips across 33 prompts spanning Western and Global South music styles
• Human: lewtun/music_genres, real recordings across multiple genres

Global South styles included: afrobeats, highlife, mbalax, bongo flava, benga, tizita, amapiano, mbira, cumbia, baile funk, reggaeton, andean, salsa, bhangra, carnatic, baul, dangdut, mor lam, gnawa, sha'bi, dancehall, zouk, Hawaiian slack key.

Results

• F1 Score: 1.000

• Accuracy: 1.000

• Confusion matrix:

• [[ 7 0]

• [ 0 14]]

• Top-left = correctly identified human

• Top-right = human wrongly called AI (false positive)

• Bot-left = AI wrongly called human (false negative)

• Bot-right = correctly identified AI

Known limitations

• Trained on MusicGen-small outputs only. Suno, Udio, and other generators not yet validated.
• Global South prompts were generated by a Western-trained model and may not fully represent those musical traditions.
• MP3 compression, pitch shifting, and time-stretching degrade accuracy.
• A high human score is not legal proof of human authorship.

Use it

from huggingface_hub import hf_hub_download
import torch
import torch.nn as nn
from torchvision.models import mobilenet_v3_small

model = mobilenet_v3_small(weights=None)
model.classifier[-1] = nn.Linear(model.classifier[-1].in_features, 2)

weights = hf_hub_download(
    repo_id="OrpheaLabs/orphea-detect-v1",
    filename="orphea_detect_v1.pth"
)
model.load_state_dict(torch.load(weights, map_location="cpu"))
model.eval()

Links

• Artist tool: huggingface.co/spaces/OrpheaLabs/orphea-detect
• Dataset: huggingface.co/datasets/OrpheaLabs/orphea-music-detection-v1
• GitHub: github.com/ElleFife/orphea-detect

License

MIT — Orphea Labs