Guides and technical articles on AI music forensics, artifact removal, and production.
Three experiments that looked like failures — a leaked udio split, an ablation that should not have worked, a benchmark we thought we had frozen — reversed how we read ArtifactNet's strengths and blind spots. Evaluation leakage inflated scores on some generators and hid collapse on others; fixing the split changed the headline story. Why we publish negative results and the methodology mistake that cost us weeks.
ArtifactNet is 4M parameters, but wall-clock latency on long files still ruled out live forensic checks — small on paper is not automatically fast in production. Distillation, batching, and TensorRT FP16 each moved the numbers in ways the architecture slide did not predict. What shipped in v9.7.1-batch-lw, what got faster, and what is still not real-time.
Listening cues, spectrogram analysis, and automated forensic tools for identifying AI-generated music from Suno, Udio, Stable Audio, and more. Includes benchmark comparison.
Step-by-step guide to removing the metallic, watery, swirly artifacts from AI-generated music. Covers RVQ ghosting, codec residue, and HF aliasing — what they are and how to fix them.
V10 pushed SONICS zero-shot detection to ~100%, but Suno v4 recall collapsed the same week — optimizing one leaderboard hid a real-world regression. We built ArtifactBench (6,183 tracks, 22 generators) to measure codec-aware detection without saturation blind spots. Why chasing benchmark F1 stopped telling us whether ArtifactNet actually worked in production.