Truthring and Hiya
Hiya screens a call while it is happening. Truthring analyses a recording afterwards. On published disclosure Hiya is ahead of us on almost every measure that matters, and on the one axis where we differ it is a difference of purpose rather than quality.
We make one of these two products, so treat this as an interested party’s reading and check it yourself. Everything below was verified against Hiya’s own live pages on 31 August 2026, with the URL beside each finding.
Prevention beats explanation, and Hiya does prevention. Their product analyses voice characteristics during live phone calls and returns a verification in fractions of a second. A fraud stopped mid-call costs nobody anything. A recording analysed afterwards explains what happened to money that has already moved. If your problem is calls arriving now, that is the more valuable half of this comparison and it is not ours.
What Hiya publishes that we do not
This section comes first because it is longer than the one after it. Hiya is the most disclosed detector of the ten we examined, and pretending otherwise would discredit everything else on this site.
Accuracy broken out across fourteen public datasets
Average EER 2.113%, pooled 2.324%, with individual dataset EERs ranging from 0.000% to 12.099% — across ASVspoof 2013 / 2019 / 2021 / 2024, ADD 2022 and 2023, In-The-Wild, LibriSeVOC, DFADD, SONAR, Fake or Real and CodecFake. Publishing the range, including the dataset where performance was worst, is rarer than publishing an average.
blog.hiya.com · checked 31 August 2026
A named false-positive trigger
No other vendor we checked names a specific condition that produces a wrong answer. Hiya does: “pre-recorded phone tree systems and automated messaging systems — like those used for school announcements or medical appointment reminders — will trigger the digital voice warning,” and “occasional ‘false positives’ may occur.”
hiyaphone.zendesk.com · checked 31 August 2026
Model versions, published and on the result
A public model index lists authenticity digital v3 and phone v9. The API result carries a required model field, for example "phone/v1". Truthring commits to the same thing and has not shipped it.
developer.hiya.com · checked 31 August 2026
A unique identifier on every verification
Each verification carries a required id — “The UUID of the authenticity verification. Is unique” — plus a handle and an alias. This is the audit trail we argue detectors should have.
developer.hiya.com · checked 31 August 2026
Separate models for separate channels
A digital model for 16 kHz and above, and a phone model that is “the only model that supports audios of 8 KHz.” Rather than quoting one number across conditions that behave differently, they built different models for them.
developer.hiya.com · checked 31 August 2026
Certifications, and no cloning business
ISO 27001 certified by A-lign, an accredited body, announced 2 May 2022, plus SOC 2 Type 2. Truthring holds neither. Hiya also does not sell voice cloning or synthesis — the only vendor-built detector we checked that has no product on the other side of the problem.
hiya.com/newsroom · checked 31 August 2026
Side by side, on what each publishes
| Published item | Truthring | Hiya · checked 31 August 2026 |
|---|---|---|
| False positive rate | Committed, framework published. No measured figure yet. | No numeric rate found at hiya.com/products/protect/ai-voice-detection. EER published per dataset, which is a related but different measure. A false-positive trigger is named in the consumer docs. |
| Accuracy by condition | Committed by condition. [VERIFY: not measured] | Published as a 14-dataset EER breakout with the full range, not a single figure. |
| Written methodology | Published | Published in a blog post naming the benchmark, datasets and model scale. Not found as a product-page document. |
| Stated failure conditions | Published | Published as hard operating limits — 5 minutes, 64 MiB, 750 ms minimum voice — plus the named phone-tree trigger. A narrative limitations page was not found. |
| Generator coverage | 27 systems named | Not found on the pages checked. Described generically; generators implied only through the named evaluation datasets. |
| Model version on a result | Committed. Not shipped. | Published — required field on every result. |
| Reference code on a result | Committed. Not shipped. | Published — required unique UUID. |
| Named people | One, honestly | Leadership named. Individual researchers not named; the benchmark authors credited are the benchmark’s, not Hiya’s. |
| Certifications | None. Stated plainly. | ISO 27001 and SOC 2 Type 2. |
| Audio retention | Committed. [VERIFY: window not yet fixed] | No audio-specific period found at hiya.com/legal. General necessity clause only. Trust centre states data is not used to train their models. |
Where a row says not found, it means we checked that URL on 31 August 2026 and did not find it — not that Hiya does not publish it somewhere we did not look. Corrections: corrections@aivoicedetctor.com
The one thing worth reading carefully
Hiya’s headline reads: “Over 99% accuracy vs. the most complex ‘In-the-wild’ datasets, setting the industry benchmark.” The dataset is named and the link resolves to Fraunhofer AISEC, which is far better sourcing than an unattributed percentage.
One distinction a careful reader should draw anyway: Hiya is not listed on the Fraunhofer page. The dataset is public and anyone can evaluate against it, so this is a vendor self-report measured on a third-party dataset — not a third-party evaluation of Hiya. That is a meaningful step above an unsourced claim and a meaningful step below an independent audit, and it is worth knowing which one you are reading. Truthring will be in exactly the same position when it publishes its first numbers, and we will say so there too.
Separately: the 97.4% figure on Hiya’s homepage is a customer testimonial about spam and fraud call blocking, not about deepfake detection. The two should not be conflated, and we mention it only because the numbers sit near each other on the page.
Where Hiya is the better choice
- You need the fraud stopped, not explained. Their detection runs during the call. Ours runs after it. For a contact centre, a carrier or anyone whose exposure is live, that difference is the whole decision.
- You are an individual who wants protection on your phone. Hiya ships consumer apps with over ten million downloads and a browser extension. Truthring will be a service you submit a file to. Those are not the same product and most people need theirs, not ours.
- You need a vendor that has passed an audit. ISO 27001 and SOC 2 Type 2 against our none. For a regulated buyer that is not a preference, it is a gate.
- You want a supplier with no stake on the other side. Hiya sells no cloning or synthesis product. Neither do we, but they have been that way at scale for longer.
- They absorbed the specialist. Hiya acquired Loccus.ai in July 2024 and the technology became Hiya AI Voice Detection. If you were evaluating Loccus, this is where it went — the full disposition.
Where we differ, honestly
Two things, and neither is a claim to be better.
- We publish the false positive rate as a rate. Hiya publishes EER, which folds both error directions into one crossover point. EER is a legitimate and more rigorous measure than most of this field uses. It also does not tell a person being accused how often genuine speech gets flagged at the threshold actually deployed — and Hiya says explicitly that thresholds are the customer’s to choose. That gap is the one we are built to fill.
- We are aimed at the person on the receiving end. Hiya sells to carriers and enterprises who deploy detection across traffic. We are aimed at someone holding one recording that matters — a journalist, a lawyer, a parent, an investigator — who needs a result that a third party can check. Different buyer, different artefact.
Questions
Is Hiya or Truthring more accurate?
Nobody can answer that from published information. Hiya publishes EER across fourteen public datasets; Truthring has published no measured figures at all yet. Comparing a rigorous number against no number is not a comparison. When our results exist they will be measured on our own held-out data, which is not the same test set, so the two still will not be directly comparable.
What is the actual difference between the two products?
Timing. Hiya analyses voice during a live call and returns a verification in fractions of a second. Truthring analyses a recording after the fact and returns a report a third party can examine. Prevention and forensics are different jobs and most organisations eventually need both.
Hiya publishes more than you do. Why would anyone choose Truthring?
Today, for a live-call problem, they probably should not. Where we differ is the artefact: a per-clip report carrying the method version, the false positive rate for that audio condition, and a reference code, aimed at someone who has to defend a conclusion later. That is a different need from screening call traffic at scale.
What happened to Loccus.ai?
Hiya acquired it in July 2024 and rebranded the technology as Hiya AI Voice Detection. Every loccus.ai URL now redirects to Hiya. If you were comparing Loccus, this page is the successor.
Should I trust a comparison written by a competitor?
No — verify it. Every claim here names the page we checked and the date. Hiya is the vendor this comparison treats most favourably, which should tell you the page was written by reading their documentation rather than by deciding the conclusion first.
All claims about Hiya verified against their live pages on 31 August 2026. Sources: hiya.com/products/protect/ai-voice-detection · developer.hiya.com/docs/audio-intel · blog.hiya.com · hiya.com/newsroom · hiya.com/company/about · hiyaphone.zendesk.com. Re-checked quarterly; corrections published with a date rather than made silently.