Is This Song AI Generated?
Free AI Music Detector

AI-generated music has measurable acoustic differences from human-recorded music: unusual energy distribution above 8 kHz (the 'metallic sheen' from AI vocoders), too-consistent timing (no human micro-timing variations), low spectral flatness variance (AI character changes less over time), and lower dynamic range. Drop your song into our free detector and we'll measure all 5 signals in your browser.

Honest answer: no AI music detector is 100% accurate. State-of-the-art ML classifiers reach 85-95% on clean test sets but degrade on real-world audio. Our statistical detector is in the 70-85% accuracy range — useful as a screening signal, not a verdict. The tool shows you the breakdown of each signal so you can interpret rather than just trust a single number. False positives (real music flagged as AI) are most common with heavily auto-tuned vocal tracks and extremely compressed modern pop.

The detector picks up acoustic artifacts common to most current AI music generators: Suno (all versions), Udio, Mureka, MusicGen, ElevenLabs Music, MusicFX, and similar. The signals it measures (high-freq spectral envelope, timing consistency, dynamic range) are characteristic of AI vocoder output regardless of which model produced it. It's less reliable on tools that use diffusion-only audio synthesis with no vocoder step.

Common false-positive triggers on human-recorded music: heavy auto-tune (compresses pitch variance similarly to AI vocals), modern pop production (heavily compressed dynamic range), 100% MIDI-programmed instruments (no human timing variance), and certain electronic genres (EDM with sample-based percussion looks like AI to the timing-consistency signal). Look at the per-signal breakdown — if only the timing signal flagged but the rest were clean, it's likely a false positive.

No. The audio file stays on your device. Decoding is done by the Web Audio API. All 5 signal analyses run locally with JavaScript DSP — no upload, no server processing, no log. The only network call is to load the page itself.

AI music generators use neural vocoders that decode audio from a learned latent space. Vocoders introduce specific high-frequency artifacts — typically a 'metallic sheen' between 8 kHz and 16 kHz that has unusual harmonic spacing not found in natural recordings. Our signal #1 measures this directly by comparing the spectral envelope above 8 kHz to expected human-recording statistics.

Signal 1: High-frequency artifact density — unusual energy patterns above 8 kHz from AI vocoder output. Signal 2: Spectral flatness variance — does the spectral character change naturally over time, or is it suspiciously constant? Signal 3: Onset timing consistency — do note onsets have natural human micro-timing variation, or are they too regular? Signal 4: Harmonic-to-noise ratio in vocal range — AI vocals have specific HNR signatures. Signal 5: Crest factor / dynamic range — AI mixes are often more compressed than human mixes.

Mostly yes. Mastering changes loudness, EQ, and stereo width but doesn't remove the underlying AI vocoder fingerprint above 8 kHz or the unnaturally consistent timing. Mastering may slightly compress the dynamic-range signal, but the other 4 signals remain. Heavy mastering with aggressive saturation and compression can produce false negatives in some cases.

No. This is a screening tool with statistical accuracy in the 70-85% range. For legal use cases (copyright disputes, distribution disputes, content authenticity claims) you need a forensic-grade audio analysis from a qualified expert. Our tool is appropriate for: deciding which tracks to investigate further, screening submissions, satisfying personal curiosity. Not appropriate for: courts, contract enforcement, official disclosures.

About 5-15 seconds for a typical 3-minute song. The pipeline decodes the audio, computes a spectrogram, runs all 5 signal analyses in parallel, and presents results. There's no model download, no server round-trip, and no warmup — just JavaScript DSP. Larger files take proportionally longer (a 10-minute song may take ~30 seconds).

Partially. 4 of the 5 signals work on any audio (high-freq artifacts, spectral flatness, timing consistency, dynamic range). The harmonic-to-noise ratio signal is most informative on vocal tracks; on instrumentals it's still computed but contributes less weight. Result: instrumental analysis is slightly less accurate but still useful.

Look at the per-signal breakdown the tool shows you. If only one signal flagged but the others were clean, the verdict is weak. If 4 of 5 signals flagged, the verdict is strong. This tool is designed to be transparent — every measurement is shown to you. Don't trust a single number; trust the pattern across signals. And remember: no detector is 100% accurate.