@audio/measure-lossy

import lossy from '@audio/measure-lossy'

Is this "lossless" file really a lossless master, or an upsampled/transcoded MP3, AAC, Vorbis or Opus source relabelled as WAV/FLAC? Reads the file's long-term average spectrum (Welch, via @audio/spectral-ltas) plus a per-frame max-hold spectrum, finds where content stops (a sustained dB drop, not a notch), measures how steep that edge is, and counts the "swiss cheese" spectral holes and the MP3-specific ~16 kHz scalefactor-band notch that give away a lossy source. This is the numeric form of the manual "look at the spectrogram" method popularised by Spek and aucdtect/"Fakin' the Funk" guessers — the same heuristics (spectrogram cutoff, sfb21 notch), just measured instead of eyeballed.

let r = lossy(data, { fs: 44100 })
r.cutoff      // Hz — estimated content bandwidth
r.lossy       // boolean verdict
r.confidence  // 0..1, heuristic — not a calibrated probability
r.source      // best guess: 'mp3-128', 'opus-96', 'lossless', 'upsampled', 'unknown', …
r.evidence    // the numbers behind the verdict (see below)
r.spectrum    // { freqs, db } — the LTAS used, for plotting

data is a mono Float32Array, or Float32Array[] (one per channel — mixed down internally).

evidence:

Method (cited in full in lossy.js): LTAS + max-hold spectra over the analysed span; the highest frequency where the (smoothed) LTAS drops minDrop dB relative to the band 2 kHz below and stays down to Nyquist is the cutoff — max-hold arbitrates, extending the estimate toward Nyquist when transients demonstrably punch past what looked like a wall, but never resetting a genuine low cutoff back up on one loud click. Spectral holes and the MP3 sfb21 notch follow Hennequin et al. ("Codec-independent lossy audio compression detection", ICASSP 2017) and Yang et al. ("Detecting MP3 compression history", 2008); the bitrate→cutoff source table is LAME's/aoTuVb's/Opus's published defaults, with citations and measured caveats in lossy.js.

This method loses power above roughly 128-160 kbps — that's a documented, known limit of spectral-cutoff detection generally (aucdtect and Spek-based "guessers" share it), not specific to this implementation: well-tuned psychoacoustic encoders at moderate-to-high bitrates can be spectrally near-transparent. It is also fooled the other way by naturally band-limited material — a cassette dub, a narrow-band mic, an aggressive de-esser, or plain speech (see the audio-lena/raw test in test.js) can read exactly like a lossy cutoff. evidence exposes every number precisely so a caller/UI can judge instead of trusting a single boolean; confidence is a heuristic score, not a calibrated probability.

Also exported as an audio.js stat manifest (./audio — a.stat('lossy'), no pre-fold needed since the kernel already accepts multichannel input).

Use when: verifying a purchased/downloaded "lossless" file is what it claims to be; auditing a library for mislabelled transcodes; flagging upsampled masters before mastering/distribution.

Part of @audio/measure — the measure family umbrella.

MIT © audiojs