Can you trust what you hear?
Truthring examines a recording and reports whether the speech in it was generated by a machine — with a stated confidence, the conditions that shaped the result, and a reference code someone else can check. It is built for the audio people actually receive: voice notes, call recordings, voicemail.
Truthring is pre-launch. The analysis pipeline is described in full on the methodology page, and measured results will be published on accuracy before the product opens.
Consistent with a rendered file rather than a captured one.
Voice stopped being proof of identity
For as long as telephones have existed, hearing someone’s voice has been treated as evidence that it was them. That assumption held because reproducing a specific person’s voice convincingly took equipment, skill and time. It now takes a subscription and about a minute of their recorded speech — from a podcast, a conference talk, a video on a family account, a voicemail greeting.
What has not caught up is everything built on the old assumption: the bank that confirms a transfer by phone, the finance team that takes an authorisation from a familiar voice, the court that treats a recording as self-authenticating, the parent who acts on a call from their child.
What a detector can actually look at
Speech that sounds convincing to a listener and speech that was captured are different things, and the difference is not in how good it sounds. A real recording passes through a physical chain — a room, a body, a microphone — and each stage leaves traces the generator had no reason to reproduce faithfully, because sounding right to a human was the only goal.
Recording chain
Whether the audio carries the signature of a real microphone in a real room, or of a file rendered directly to disk. Room reflection, handling noise and the noise floor of an actual sensor are difficult to fake and usually absent by default.
Generator signature
Fine regularities left behind by a particular synthesis system. This is what lets a verdict name the generator rather than say only that something was made by a machine. It decays fastest, because vendors ship new models between our retrains.
Prosody under stress
How a voice behaves at the edges: an interruption, a laugh, a restarted word, a sentence abandoned halfway. Generated speech is trained on clean read audio and is least convincing where real conversation is messiest.
The full pipeline, including preprocessing, thresholds and versioning, is on the methodology page. Which generators can be named is on the coverage page.
Three outcomes, and two of them are not opposites
Likely synthetic is the strong verdict. It is only reached when positive evidence is present — something had to be found to say it. Likely human is the weak one: it can mean the speech was human, or it can mean the evidence was destroyed by compression on the way to you. It is not a clearance.
Signals consistent with generated speech were found. Where the clip matches a signature we hold, the report names the system. Where it does not, it reads unknown generator rather than guessing at the nearest match.
No signals of synthesis were found. On a clean recording that is meaningful. On a compressed phone call it is close to no information at all, and the report says which case you are in.
The audio does not support a verdict — too short, too degraded, too much background. This is a result, not a failure, and under the intended pricing it is never billed.
The number most detectors will not publish
Detectors advertise a single headline accuracy figure. It is almost always the detection rate — how often the tool catches a fake — and it is easy to make large. Flag more aggressively and it rises.
What that conceals is the price of raising it: the rate at which genuine human speech gets flagged as fake. The two move in opposite directions. Showing only the flattering half is not exactly a lie; it shows you one side of a trade-off and lets you assume the other side is free.
It is not free, and the people who pay are the ones being accused — the support agent whose call is flagged, the witness whose voicemail is disputed, the candidate who loses a job over an artefact of compression. For them the false positive rate is the product. Truthring will publish both, broken out by audio condition, with the dataset described and the date attached.
Nothing is measured yet, and we are not going to pretend otherwise. Every accuracy figure on this site is marked pending. The evaluation framework is published in advance on accuracy, benchmark and the dataset page, so the method can be argued with before any number exists to defend.
Four conditions that defeat this
Publishing these costs us nothing we were entitled to keep, and it is the difference between a measurement and a marketing claim.
Audio re-recorded through a speaker
Playing a generated clip aloud and capturing it on a phone gives the file a genuine recording chain. It defeats the strongest signal available, and it is not difficult.
Heavy compression
A phone network discards most audio above roughly 3.4 kHz, which is where much of the evidence lives. Detection on call recordings is materially weaker than on clean files.
Very short clips
Below a few seconds of speech there is not enough material for the later passes to contribute, and confidence should be read accordingly.
A system newer than our last retrain
The verdict may still be right while attribution drops to unknown generator. Naming the generator decays faster than detecting one.
Different readers, different stakes
A parent checking a frightening voicemail and a bank reviewing a disputed authorisation need the same analysis and very different things around it. Each of these pages is honest about what detection cannot do for that reader.
Individuals and families
A call that did not sound right, and what to do in the next ten minutes.
Read moreJournalists
Verifying audio before publication, and handling a source who denies a genuine recording.
Read moreLegal teams
What a report can and cannot support when admissibility is decided by a court, not a vendor.
Read moreSecurity teams
Post-incident forensics, evidence preservation, and the employee wrongly flagged.
Read moreAll twelve audiences, including the section on who should not buy this.
What happens to a recording you send
People submit recordings of family in distress, calls from their bank, and evidence in disputes they have told nobody about. That is not ordinary file upload, and the trust pages are written to be checked rather than to reassure.
Truthring holds no SOC 2 report, no ISO 27001 certificate and no independent audit. Saying so is more useful to you than staying quiet about it, and a buyer discovering it during diligence is worse than reading it here.
Security
The path a clip takes, and what is deliberately not claimed.
ReadData retention
Audio, features, verdict and reference code have four different lifetimes.
ReadTransparency
What is published, what is not, and how corrections are handled.
ReadCompliance
Lawful use, and why no result may be the sole basis for a decision.
ReadCommon questions
What is Truthring?
Truthring is an audio verification service that examines a recording and reports whether the speech in it was generated by a machine, with a stated confidence and the conditions that shaped the result. It is a product of Lacewing Technologies, an independent software company in Navi Mumbai, India, which also operates TextSight.ai.
How accurate is it?
No measured figure exists yet, and we will not publish one until it does. The evaluation framework is published in advance on accuracy and benchmark, including the false positive rate, which most competitors omit. Any accuracy claim you see elsewhere on this site is marked pending.
Can a result be used as proof?
No. It is a probability produced by a stated method with a stated error rate. That makes it usable as one input alongside others and unusable on its own. Never make it the sole basis for accusing, disciplining, dismissing or prosecuting anyone.
Can it detect a voice clone of a specific person?
It detects whether speech was generated, not whose voice was imitated. Those are different questions, and a system answering one cannot answer the other. Speaker verification asks whether a voice belongs to a claimed person; Truthring asks whether a machine produced it.
What should I do right now if a call seems wrong?
Hang up and call the person back on a number you already have stored — not one the caller gives you. A cloned voice cannot answer a phone you dialled. Analysis explains afterwards what a recording was; the callback stops the fraud while it is happening.
Is Truthring available yet?
Not yet. The site documents the method, the evaluation framework and the limits in advance of launch, deliberately, so the approach can be argued with before there is a product to sell. You can request early access.
Check a recording when it opens
Or read the methodology first and decide whether the approach is one you would trust with a result that matters.
Reviewed