Callcheck - Audio Triage for Suspected Voice Fraud
Every tool that promises to detect a cloned voice is guessing. This one doesn’t promise it.That sounds like a strange thing to put at the top of a sales page. It is the reason to buy this one.Voice-clone detectors are trained on one generation of speech synthesis. The next generation defeats them. So does the compression that every phone call and every messenger voice note already carries. What you get is a confident-looking percentage next to an answer that is a coin flip — and somebody wires the money on the strength of it.Callcheck does two things that hold up instead.1. It measures the fileReproducible numbers, no model, no training data, nothing to go stale: • Bandwidth and the shape of the upper edge. A lossy codec cuts the spectrum off vertically. A microphone never does. Verified against real encodes: AAC at 32 kbit leaves a cliff at 12.4 kHz, MP3 at 64 kbit at 16.5 kHz, Opus at 20.0 kHz — and the uncompressed original shows no cliff at all. • Noise floor and the character of the pauses. Every microphone recording carries room tone. A floor below −85 dBFS was not produced in a room. • Splice points. Abrupt, persistent changes in room tone between segments. • Clipping, DC offset, channel layout, container metadata.2. It checks those numbers against what you were toldThis is the part that turns measurements into a finding.A telephone channel carries about 300–3400 Hz. A messenger voice note is Opus or AAC at low bitrate. A direct microphone recording has no codec edge. These are properties of the channel, not of the voice — and a channel imposes them on everything that passes through it.So you tell Callcheck what the file is supposed to be:$ callcheck sprachnachricht.wav --claim phone Angegebene Herkunft: Telefonanruf (Festnetz/Mobil) 2 Widersprüche zur Angabe [WIDERSPRUCH] Mehr Bandbreite als „Telefonanruf" zulässt gemessen 24.0 kHz, erwartet bis 4.2 kHz [WIDERSPRUCH] Keine Codec-Kante, obwohl der Kanal eine hinterlassen müsste Kantensteilheit 1 dB (eine Codec-Kante liegt typischerweise über 28 dB)That is not an opinion about whether the voice is real. It is a statement about where the file cannot have come from — and it is checkable by anyone who has the file.Seven channel profiles: telephone, HD voice, messenger voice note, voicemail, video meeting, direct recording, or none.3. It refuses to produce false alarmsA tool like this earns its place by staying quiet when nothing is wrong, so the logic was built and tested that way: • A microphone that runs to the Nyquist limit is not flagged for missing a codec edge — the sample rate is the limit, not a codec. • A single blip in the noise floor is not a splice; the change has to persist, and a file too short to confirm that reports nothing. • A splice in a phone call is a contradiction because a call is one take. The same splice in an edited recording is merely worth noting.Tested across fourteen file-and-claim combinations against synthesised audio with known properties, plus real AAC, MP3, Opus and µ-law encodes.4. It gives you the procedure that actually worksBecause none of the above settles the question, every report ends with the verification protocol — the part that stops the fraud whether or not the audio was synthetic:Do not reply on the channel the request arrived on. Whoever cloned the voice also controls the channel.Call back on a second, independent channel. Switching channel is the single most effective measure there is.Ask something only the real person could know. A cloned voice can speak but cannot remember.Treat time pressure and secrecy as what they are — not signs of urgency, but the core tool of this fraud. They exist so that you do not call back.Eight steps, in both languages, printable on its own with --playbook so it can go on the wall next to the phone in accounting.What you get • One Python file. Standard library only, Python 3.8+. Uses numpy if present for speed, and a built-in FFT if not — verified identical to fifteen digits. • Read-only, fully offline. No model download, no upload, no network. The recording never leaves the machine, which matters when it is evidence. • WAV directly; MP3, Opus, M4A, AMR and µ-law telephony via ffmpeg if installed. • German and English throughout, including the report and the protocol. • Exports: standalone HTML report for the client or the file, and JSON. • --fail-on-contradiction returns exit code 2.Who it is for • Accounting and management, for the moment a voice message asks for a transfer • IT service providers and agencies who get that call from a client • Anyone who has to write down what was examined, and what was concluded, afterwards
Get it → tracekeeper.gumroad.com