A probability is not an opinion, and neither is proof
You will know the evidential rules in your jurisdiction better than we do. What we can be precise about is what our output actually is: an automated measurement with a published method, a version number and a stated error rate. What you build on it is your judgement.
Nothing on this page is legal advice, and we are not your lawyers or your client’s.
The exposure that is new
The old question about a recording was whether it had been edited. That question had settled answers and a professional discipline behind it. The new question is whether the speech ever happened at all, and it arrives in matters that were never audio cases: a voice note in a family dispute, a recorded instruction in a commercial claim, a call in an employment tribunal, a voicemail in a harassment complaint.
It cuts both ways in the same file. A party may tender a fabricated recording. A party may also denounce a genuine one as fabricated, knowing the other side cannot easily prove a negative, and knowing that the mere availability of synthesis has made that denial cheap. The second is the harder one to meet, and it is becoming the more common.
What a report can support
A verified Truthring report is a dated record of an automated analysis. It states the SHA-256 hash of the file submitted, the timestamp, the engine version, the verdict and confidence, the audio conditions measured, and the limitations that applied to that specific clip. It carries a reference code and a permanent URL, so the opposing party can retrieve the same document rather than take your description of it.
Used properly, it can support a limited set of propositions. That the file bearing this hash was analysed on this date by this version of this engine. That the analysis found — or did not find — patterns associated with known synthesis systems, at a stated confidence. That the recording’s measured condition placed a ceiling on how much that finding is worth.
It is also useful before proceedings. A screening result is cheap and fast, and it tells you whether it is worth instructing a forensic examiner at all. It can support an application, frame a request for the original file in disclosure, or give you a reason to press for the device.
What this cannot do for you
It cannot decide admissibility. That is for the tribunal. The tests for novel or technical evidence differ between jurisdictions, between civil and criminal proceedings, and between first instance and appeal. [VERIFY: verify the applicable test in your markets] A vendor telling you its output is admissible is telling you something it has no standing to say.
It is not an expert report. There is no named examiner behind it who can be cross-examined on their reasoning, no instruction, no duty to the court. Where the matter turns on the audio, instruct a qualified forensic audio examiner and let them decide what weight, if any, an automated screen deserves within their own analysis.
It cannot authenticate. A likely human verdict means no signature we hold was found. It is the weaker verdict, it can be produced by a synthesis system newer than our training data or by a clip processed after generation, and it should never be tendered as proof that a recording is genuine.
It cannot identify a speaker. We do not do voice comparison or speaker attribution. Whether the voice belongs to your client is a different discipline with a different error profile.
It cannot date a recording or establish provenance. Codec history tells you something about what a file has passed through, not who made it or when.
It cannot be the reason a case is run or abandoned. A person exercising judgement makes that call, and the confidence figure is an input to it.
Chain of custody is your client’s, not ours
We receive a file. We do not know how it was obtained, whether it is the original or a copy, what software has touched it, or whether the person who supplied it is entitled to hold it. We analyse it, return a result, delete the audio and retain a one-way hash. That hash lets a later copy be matched to the file we saw. It does not make us a custodian, and no part of our process is a substitute for the record your side keeps.
Practical consequences, which your litigation support team will already recognise:
- Preserve the original file, on the original device where possible, before anything is uploaded, converted or forwarded. A screening check should be run on a copy.
- Hash the original yourself at the point of receipt, so the hash in our report can be tied to something you can evidence independently.
- Record how the recording was obtained. Lawful recording rules vary — some jurisdictions require one party’s consent, some all parties — and a recording obtained unlawfully may be worthless in the proceeding and a liability in itself. [VERIFY: verify the rule in your markets]
- Consider your data protection obligations before uploading a recording of a person who is not your client. Analysing a voice is processing personal data in many regimes, and you will need a basis for it.
Common questions
Will you provide a witness statement or attend as an expert?
[VERIFY: state plainly whether you offer this, and if so on what terms and with which named individual. If you do not, say so here rather than leaving it ambiguous.]
Can we run the other side’s audio without telling them?
That is a question for you, and it usually has both a professional-conduct and a data-protection dimension. Note only that a covert analysis rarely stays covert, and a result you are unwilling to disclose is a result you cannot rely on.
What error rate should we put to a witness?
The one that applies to the condition of the actual clip, not the headline figure. Rates for clean audio and for compressed telephone audio differ materially. The accuracy page breaks this down and the limitations page lists the conditions under which we say a result should not be relied on at all.
How long is a report retrievable?
[VERIFY: retention period for verified reports and their verification URLs] Take your own copy at the time; do not assume a URL will still resolve at trial.
Screen first, instruct second
Use a report to decide whether the audio is worth an expert. Do not use it as the expert.
Reviewed