0don/trashbin-plus-ai

trashbin+ AI music detector

Detects AI-generated songs from a 30 second Spotify preview. Used by the trashbin+ Spicetify extension, where it runs locally in the browser through onnxruntime-web.

Model

• detector-mn10-v5.onnx (17 MB): EfficientAT MobileNetV3 mn10_as (AudioSet pretrained, 4.2M parameters) fine-tuned as a binary AI vs human classifier.
• Input waveform: float32 [N, 320000], mono 32 kHz, 10 second windows. The log-mel frontend is inside the graph.
• Output logit: [N]. The extension scores the start, middle and end window of a preview and averages the logits, then applies a sigmoid.

Training data

About 22,000 clips from 50+ sources, all re-encoded to the Spotify preview format (96 kbps MP3, 44.1 kHz stereo, 30 s):

• AI: Suno v2 to v6, Udio, Lyria 3 and 3.5, ElevenLabs Music v1 and v2, Mureka, MiniMax, MusicGPT, Stable Audio 1 to 3, MusicGen, Riffusion/Producer, ACE-Step, YuE, DiffRhythm, SongGen, HeartMuLa, Mubert, Boomy, AIVA, Soundful, Soundraw, TemPolor, Brev, and AI artists on Spotify.
• Human: Spotify releases from before 2022, FMA, MTG-Jamendo, MusicCaps.
• Public datasets used include ArtifactBench, HAIM and SONICS (CC BY-NC 4.0), MUSIC8K (CC BY 4.0) and Echoes (CC BY-SA 4.0). Because part of the training data is non-commercial, the weights are released under CC BY-NC 4.0.

Results (v5, held-out eval set, threshold 0.8)

• AUC 0.994.
• Human songs flagged: Spotify 0 of 299, FMA 3%, MTG-Jamendo 0 to 10%, MusicCaps 0%.
• Caught: AI artists on Spotify 95%; neural generators (Suno, Lyria, ElevenLabs, Mureka, MiniMax, MusicGPT, Stable Audio, MusicGen, ACE-Step) 87 to 100%.
• Weak: AIVA 40%, Loudly releases 53%, Mubert 67%, Boomy 70%, current Udio 71%. These assemble recorded loops or render MIDI with sample libraries and leave few audio artifacts.

Scores are probabilities from one model, not proof. Training code: ai/.