Upsampling

The Upsampled finding identifies a high-rate PCM container whose decoded content has a persistent high-frequency dead zone compatible with a lower-rate source. It is an upsampling heuristic, never a proof of the file's history: a genuinely recorded or deliberately filtered low-bandwidth master can produce identical PCM samples.

Signal tested#

The detector runs only when the declared sample rate is above 48 kHz. It repeatedly analyses AAC-sized MDCT windows from the decoded stream, rather than using the displayed FFT cutoff. The MDCT view is useful here because it can test whether a high-frequency region stays empty over many time-local frames instead of treating a single faint spectral bin as meaningful content.

A result is flagged when all of these project thresholds hold:

The analyzer samples one out of four overlapping MDCT hops and caps its work at 240 frames. This bounds analysis time for long files, so the result is a measurement of the sampled programme rather than every possible transform frame.

Result and interpretation#

Upsampled means that the conditions above were observed. The detail text states the estimated content boundary and the container rate, and explicitly says that limited bandwidth is not proof of resampling. A Clean result means this test did not establish the signature; ultrasonic noise, later processing, or a source with a different resampling signature can prevent a finding.

The displayed spectral cutoff is intentionally not the verdict input. The full-track FFT is useful to inspect a recording, but it can read near Nyquist because of isolated noise even when most MDCT frames have a dead zone, and it can read low for a native dark recording. Keeping the measurement and verdict separate avoids silently turning a descriptive spectrum column into an accusation.

Limits#

No spectral rule can distinguish two files with identical samples but different provenance. Native acoustic, archival, analogue-tape and low-pass-filtered recordings can have little ultrasonic energy. Conversely, injected ultrasonic noise can obscure the signature. The numerical thresholds are project thresholds, calibrated engineering choices rather than a published classification standard or a probability of upsampling.

The detection is about PCM bandwidth. DSD has its own source check because its shaped ultrasonic noise does not follow the same model; see DSD authenticity.

Tests#

The Rust unit and integration suites create controlled spectra and exercise the classification boundary:

cargo test -p flaccompagnon-core analysis::detections
cargo test -p flaccompagnon-core analysis::analyzer

For a manual check, create a known low-pass source, resample it to 96 kHz, then import it. This validates the plumbing and displayed result, not a universal accuracy claim:

ffmpeg -n -f lavfi -i 'sine=frequency=1000:sample_rate=44100:duration=30' -af 'lowpass=f=20000,aresample=96000' -c:a pcm_s24le upsampled-fixture.wav