Skip to content
AI Music Detector

Detection methodology and analysis information

AI Music Detector runs entirely in your browser. This page documents every measurement it takes and every one it deliberately does not.

Last updated August 2026

Engine card

  • Analysis Engine: AI Music Detector
  • Type: AI Music Detector — deterministic signal measurement, not a trained neural classifier
  • Training data: none — no data was used to fit parameters; weights were set from published acoustic reasoning and are stated openly below
  • Execution: entirely client-side, Web Audio API + in-page FFT
  • Last evaluation: no formal held-out evaluation completed
  • Published metrics: none, by design (see accuracy)

Inputs and preprocessing

Supported containers: MP3, WAV, FLAC, AAC, M4A, OGG and WebM audio, up to 25 MB. Minimum duration 10 seconds. Decoding uses the browser’s native audio engine; anything it cannot decode is rejected rather than guessed at. Channels are summed to mono for spectral analysis, with the stereo pair retained for correlation.

Between two and five analysis windows of four to eight seconds are selected at evenly spaced positions across the file. Windows whose RMS falls below 0.0015 are discarded as silence. If fewer than two usable windows remain, the analysis stops with an explanatory error rather than producing a low-quality guess.

Features and weights

Spectral ceiling — highest FFT bin holding more than 0.35% of the peak bin magnitudeweight 1.1
Segment agreement — mean pairwise RMS distance between segment log spectraweight 1.2
Timbral movement — normalised standard deviation of the spectral centroid across framesweight 0.9
Dynamic range — crest factor in dB (peak over RMS)weight 0.7
High-band energy — share of spectral energy above 14 kHzweight 0.6
Stereo correlation — inter-channel correlation across the fileweight 0.5

Spectral flatness is measured and displayed for context but is not currently weighted, because we have no defensible basis for a direction.

Scoring and calibration policy

Each feature maps to a sub-score between roughly 0.35 and 0.7. The weighted mean is then shrunk toward 0.5 by a factor of 0.85 and clamped to the 15–85% interval. A margin of ±10 or ±15 percentage points, depending on confidence, is reported as a plausible range. There is no random component anywhere in the pipeline: the same file always produces the same result.

Classification rules

  • Inconclusive when confidence is Low, the score lies between 42% and 58%, or segment scores span more than 18 points
  • Likely AI-generated at 70% or above with at least Moderate confidence
  • Possibly AI-generated from 58% to 69%
  • Likely human-created below 42%

Confidence never exceeds Moderate in this version. High confidence is reserved for a future model with published evaluation data behind it.

What this model deliberately does not do

  • Generator attribution. It will never say a track came from a named product. Attribution requires per-generator evaluation we have not performed.
  • Stem or vocal analysis. No source separation is applied.
  • Watermark or Content Credentials reading. Planned, not implemented.
  • Fingerprint lookup. No database of known tracks is consulted.

Roadmap

The result contract already carries fields for embedding-based classification, generator signatures, provenance signals and per-stem results, so those can be added without changing the interface. In priority order: a documented evaluation dataset; a learned classifier over audio embeddings; watermark and Content Credentials detection; robustness testing against recompression and resampling.

Version history

  • Initial public release — Six weighted DSP features, conservative calibration, first-class inconclusive state, fully client-side.

Return to the detector or read how the pipeline runs step by step.