A lossy file renamed to .flac looks lossless, weighs lossless, and fools every player you own. It can't fool the arithmetic the encoder left behind — and this reads it for you.
MP3, AAC, Vorbis and — partially — Opus, through three independent kinds of evidence. A conviction needs two of them to agree.
Find the fake FLACs in your music library.
A lossy codec throws away the top of the spectrum and never gives it back. FLAC Detective reads each file, spots the fingerprints that loss leaves behind, and tells you which files are real and which are fakes.
pip install flac-detective # needs Python 3.10+
flac-detective /path/to/music # scan a file or a whole folderEvery file gets a verdict, like a traffic light:
✅ AUTHENTIC real lossless → keep it
❓ WARNING borderline → give it a listen
⚠️ SUSPICIOUS probably a transcode → likely a fake
❌ FAKE_CERTAIN definitely a fake → replace it
The scan only reads your files — it never changes anything.
🟢 New to all this? → Start Here — the 5-minute beginner's guide No command line, no jargon. From "what is this?" to "I checked my music".
Add --format html and you get a single self-contained page: a triage table sorted
worst-first, plus a spectrum plot for every flagged file. The MP3 "cliff" — a sharp drop
well below the real ceiling — is right there for the eye, with the detected cutoff marked.
Three transcodes at different MP3 bitrates show the wall falling at different frequencies (96 kbps ~11 kHz, 128 kbps ~16 kHz, 160 kbps ~17.5 kHz); the authentic file runs full-range.
An MP3 re-saved as FLAC is lossless as a container, but the audio already passed through a lossy codec — and that leaves fingerprints. The clearest is the spectral cliff: MP3 discards everything above a bitrate-dependent frequency, so the spectrum falls off a wall where a real recording keeps going.
FLAC Detective scores each file with 12 heuristic rules built around that idea (cutoff frequency, MP3-bitrate signatures, silent-passage energy) plus protection rules so genuine vinyl rips, cassette transfers and naturally quiet recordings aren't flagged. An optional CNN sharpens borderline verdicts, and a frame-alignment rule reads the fingerprint MDCT quantisation leaves behind — the one signal that survives at 256–320 kbps, where the spectral cliff has nothing left to show. The rules sum to a 0–150 score:
| Verdict | Score | What to do |
|---|---|---|
| ✅ AUTHENTIC | ≤ 30 | keep it |
| ❓ WARNING | 31–54 | borderline — check manually |
| 55–85 | likely a transcode | |
| ❌ FAKE_CERTAIN | ≥ 86 | definitely transcoded |
The guiding principle is "protect authentic files first": a false alarm on real music is worse than missing a borderline fake. Treat AUTHENTIC as "no evidence of transcoding", not a guarantee.
What the numbers describe. Every accuracy figure this project publishes was measured on direct transcodes — lossy audio re-encoded straight to FLAC, nothing in between. Transcodes that passed through a mastering chain before reaching a disc (levels, limiting, press — what wild compilations actually sell) are a different population: on 34 owner-attested wild MP3-sourced tracks, v1.11.4 signaled 8.8 %, v1.12.0 — after repairing four admission gates that had been calibrated on lab material — signaled 50.0 %, and v1.13.0 — after feeding the depth instrument to the cells the container check consults — signals 70.6 % and convicts its first two, both owner-attested fakes corroborated by independent evidence families. False convictions remain zero across every version and population, which is the half of the design that held throughout. Every number here was pre-registered before its measurement and scored on hash-verified bytes; the remaining gap is real, measured, and being worked on rather than blurred.
→ Every rule explained: Technical Details.
flac-detective /path/to/music # scan a folder
flac-detective # interactive (prompts for a path)
flac-detective /music --format csv -o triage.csv # spreadsheet, worst-first
flac-detective /music --format html -o report.html # visual report (see above)
flac-detective /music --deep # catch high-bitrate AAC/Opus/Vorbis (slower)
flac-detective /music --advanced # show the plumbing: scores, cutoff, per-rule detailBy default the output is easy mode — a plain-language verdict and a recommended action
per file ("Almost certainly a fake — the sound stops dead at ~16 kHz, the wall of a
128 kbps MP3. → Replace it."). Add --advanced for the scores, cutoff frequencies and
per-rule reasoning. The desktop GUI has the same Advanced toggle.
Analyses FLAC, WAV, ALAC (.m4a) and APE (.ape) — codec-agnostic, and a lossy .m4a
is correctly rejected (the real codec is probed, never trusted by extension).
→ Full guide & every flag: User Guide.
pip install "flac-detective[gui]"
flac-detective-guiA desktop app (PySide6): choose a folder or drag it in, watch the progress bar, then triage a sortable, colour-coded verdict table — click any file to see its spectrum with the detected cutoff marked and the reasons for its verdict. Export to HTML/CSV/JSON.
Beyond lossy-as-lossless, FLAC Detective flags files sold as high-resolution that
aren't: 44.1/48 kHz upsampled to 96/192 kHz (a hard spectral cliff with digital
silence above it), and 16-bit audio padded into a 24-bit container. This is a
separate axis from the transcode verdict — reported as hires_verdict
(GENUINE_HIRES / UPSAMPLED / PADDED_DEPTH / …) in the CSV report, the GUI and the
Python API. A genuine 96 kHz recording that simply rolls off early reads GENUINE_HIRES,
not a false alarm.
Install options & upgrading
pip install flac-detective # base
pip install "flac-detective[ml]" # + optional CNN (Rule 12)
pip install "flac-detective[gui]" # + desktop GUI (flac-detective-gui)
docker pull ghcr.io/guillain-rdcde/flac_detective:latest # or Docker (amd64 + arm64)pip install does not upgrade an existing install — use -U to get the latest release:
pip install -U flac-detective
flac-detective --versionUse it from Python or beets
Python API:
from flac_detective import FLACAnalyzer
result = FLACAnalyzer().analyze_file("song.flac")
print(result["verdict"]) # AUTHENTIC, WARNING, SUSPICIOUS, or FAKE_CERTAINbeets plugin — triage transcodes without leaving your workflow:
pip install "flac-detective[beets]"
# in config.yaml: plugins: flacdetective
beet flacdetective # analyse & tag the whole library
beet ls flacdetective_verdict:FAKE_CERTAIN # list the certain fakesStores flacdetective_verdict and flacdetective_score on each item; an optional
auto: yes analyses files as they're imported.
Rule 12's model went through a real R&D saga, written up as a learning resource: a false-positive audit over 11 234 real FLACs, four instructive dead-ends, a debunked "AUC 0.99" caught by cross-validation, and a twist where a "fundamental limit" turned out to be an artifact of listening in mono — fixed by going stereo. Real-world specificity on 11 234 authentic FLACs climbed from 80 % to 95 %.
📖 Read the ML detective story →
- 📖 Full documentation site — getting started, user guide, technical details, API
- 🚀 Getting Started — install, first analysis, accuracy & file-safety notes
- 📋 Changelog · 🤝 Contributing · 🔒 Security
- 💬 Issues · Discussions
Licensed under the MIT License.