Forensic audio encyclopedia

An energy-weighted centre frequency, often heard as brightness.

Centroid follows arrangement, instrumentation and mix balance. Watching it move can describe contrast between passages, but no movement pattern uniquely identifies generation.

How this tool uses it: The public classification provider's internal feature weights are not disclosed. This concept is educational context, not a confirmed input to the live verdict.

A ratio describing whether a spectrum is more noise-like or tone-like.

Flatness is useful for comparing texture across regions. Noise reduction, distortion and source material can change it dramatically, so an origin claim needs much more evidence.

How this tool uses it: The public classification provider's internal feature weights are not disclosed. This concept is educational context, not a confirmed input to the live verdict.

A frequency above which little energy remains.

A visible ceiling is frequently an encoding or sample-rate boundary. It may describe the copy being inspected rather than the original production.

How this tool uses it: The public classification provider's internal feature weights are not disclosed. This concept is educational context, not a confirmed input to the live verdict.

The difference between peak and average signal level.

Crest factor describes transient headroom. It is mainly a mixing and mastering observation, and a low value is common in human-made commercial releases.

How this tool uses it: The public classification provider's internal feature weights are not disclosed. This concept is educational context, not a confirmed input to the live verdict.

The span between quieter and louder meaningful passages.

Macro-dynamics can reveal arrangement and mastering choices. They vary strongly by genre and delivery format, which makes them poor standalone provenance evidence.

How this tool uses it: The public classification provider's internal feature weights are not disclosed. This concept is educational context, not a confirmed input to the live verdict.

How mid, side, left and right energy relate over time.

Stereo analysis helps diagnose width and mono compatibility. Modern production can construct any of these relationships after recording or generation.

How this tool uses it: The public classification provider's internal feature weights are not disclosed. This concept is educational context, not a confirmed input to the live verdict.

How consistently the two channels reinforce or cancel one another.

Correlation can reveal polarity and translation problems. It cannot distinguish a microphone capture from a carefully built stereo mix with enough reliability to prove origin.

How this tool uses it: The public classification provider's internal feature weights are not disclosed. This concept is educational context, not a confirmed input to the live verdict.

How similar returned readings are across successive regions.

The provider may return values for windows across a track. Agreement can help describe the stability of its own output, while disagreement may simply follow an arrangement change or edit.

How this tool uses it: The application displays provider-returned windows but does not recompute or use them to override the provider's primary verdict.

How frequently sharp attack events occur.

Transient counts can describe rhythm and articulation. Quantised human production, drum replacement and mastering all alter the same measurement.

How this tool uses it: The public classification provider's internal feature weights are not disclosed. This concept is educational context, not a confirmed input to the live verdict.

Small deviations from a regular rhythmic grid.

Timing variation can separate a free performance from a fixed grid in favourable material. It does not separate human sequencing from generative authorship.

How this tool uses it: The public classification provider's internal feature weights are not disclosed. This concept is educational context, not a confirmed input to the live verdict.

Changes introduced by lossy formats and repeated transcoding.

Codec traces can explain why a copy differs from a master. They document a file's transmission history more readily than its creative origin.

How this tool uses it: The public classification provider's internal feature weights are not disclosed. This concept is educational context, not a confirmed input to the live verdict.

Tags, drafts, sessions, records and testimony that describe a production history.

Project files, dated drafts and contributor records answer process questions directly. Metadata can corroborate them, although tags alone are editable and often stripped.

How this tool uses it: The public detector classifies audio. It does not inspect metadata or documentary provenance.

A deliberately embedded marker readable by a compatible system.

A valid watermark can identify participation in its issuing system. Absence proves little because schemes are proprietary, coverage is incomplete and processing can damage a marker.

How this tool uses it: The application does not scan for watermarks.

A standard way to describe frequency content in short slices of audio.

A Fourier transform supports many audio tools, including spectrograms and meters. Its use is not evidence that the current provider relies on a particular spectral feature.

How this tool uses it: The public classification provider's internal feature weights are not disclosed. This concept is educational context, not a confirmed input to the live verdict.

Combining outputs from more than one model or feature family.

Ensembles can improve robustness or repeat shared blind spots. This site does not claim that its current provider uses one.

How this tool uses it: The public classification provider's internal feature weights are not disclosed. This concept is educational context, not a confirmed input to the live verdict.

Separate provider fields that should not be collapsed into one score.

Vocal and instrumental percentages describe separate provider indicators. Confidence is reported as Low, Medium or High. Neither field is a measured probability that the site is correct.

How this tool uses it: The application displays these fields separately and does not calculate a blended overall AI percentage.