{{ clsLabel }} {{ clsReason }} No measurement here is an AI-probability, and no single measurement indicates how a recording was made.

Ready to save: {{ savedName }}Save fileIf the download did not start, use this link.

Recordings

File	Origin	Length	Format	Loudness	Status
	{{ r.origin }}	{{ r.dur }}	{{ r.fmt }}	{{ r.lufs }}	{{ r.status }}

Measurements

Spectrogram · 40 Hz – {{ dFhi }} log scale · 0.5 s columns · −100 to −10 dBFS · ticks above mark time-frequency discontinuities{{ dDurText }}

Over time

Inter-onset intervals

0–1000 ms, 10 ms bins · {{ dOnsets }} onsets

Hover a measurement for its algorithm. Full definitions are under .

Known provenance

Available after the analysis finishes. The analyzer never asks how a recording was made before measuring it.

Recorded for research after the measurements were calculated, stored separately (keyed by file SHA-256), and never passed to the analysis or to a model. Changing it does not change any result above.

Origin

Generator, artist or source

Notes

Research dataset

Available after the analysis finishes.

Adds this recording’s measurements to the local dataset. The lineage names the source recording; every version with the same lineage is kept in the same split.

Change over time

B − A at the same musical moment (B series shifted by the alignment offset)

Measurement differences

Measurement	Unit	A	{{ h.name }}	Δ
{{ r.label }}	{{ r.unit }}	{{ r.a }}	{{ c.v }}	{{ c.d }}

Category	Truth	Train	Calibration	Test	Total
{{ m.label }}	{{ m.truth }}	{{ m.train }}	{{ m.calibration }}	{{ m.test }}	{{ m.total }}

{{ leakText }} Splits are assigned by a hash of the lineage (60 / 20 / 20). No classifier is trained here; the test split is for evaluation only.

Model evaluation

Confusion matrix, precision, recall, F1 and false-positive rate are calculated on the test split once a trained classifier is installed and has scored those entries. The calibration split is reserved for choosing a threshold. Nothing is shown until then.

Entries

No entries yet. Open an analyzed recording and use “Add to dataset”.

File	Category	Lineage	Split	Length
{{ e.name }}	{{ e.cat }}	{{ e.lineage }}	{{ e.split }}	{{ e.dur }}

Status

{{ clsReason }} The Lab does not turn measurements into a probability. When a model is installed, each result is shown with the model name, version, threshold, the input duration actually scored and the model’s documented limitations.

Phase 2 candidates (not installed)

{{ c.name }} {{ c.source }} Licence · {{ c.license }} Input · {{ c.input }} Browser inference · {{ c.runtime }} Recent Suno output · {{ c.suno }} {{ c.status }}

Integration contract

A model is registered with registerClassifier() in forensics-engine.js. Registration is rejected unless the descriptor carries id, name, version, licence, source, threshold, input seconds, sample rate, limitations and a run() that returns the model’s own output. Inference runs on the existing worker. Models that cannot run in the browser stay unregistered; an optional local or server endpoint can be added behind the same descriptor, with the endpoint shown next to every result. Scores are stored with the model id and version next to, not inside, the measurement report. Provenance is never an input.

All measurements use decoded samples of the whole file. Spectral measures use the mid signal (L+R)/2, 2048-sample Hann frames with a 1024-sample hop, amplitude-normalised so a full-scale sine reads 0 dBFS. Time series use 0.5 s windows. Onset detection, BS.1770 loudness, grid fitting and alignment are the CM Humanizer engine’s own functions. Measurement version {{ fxVersion }}.

{{ pageTitle }}