Truthring vs Pindrop
Pindrop publishes a false positive rate. Their Pindrop Pulse page carries “<1% False positive rate”, footnoted as “99.4% accuracy with less than 1% false positives” (pindrop.com/product/pindrop-pulse/, checked 31 August 2026). That disclosure is the thing this site argues detectors should make, and a competitor making it belongs in the first paragraph rather than a footnote.
So the honest distinction between us is not who discloses more. It is category. Pindrop scores audio inside a live call; Truthring analyses a recording after the event. For a bank’s contact centre, Pindrop is very likely the right answer and we are not one.
Pindrop publishes several things Truthring does not yet have, including the number our whole argument is built around. What follows is sourced and dated, with one substantive criticism that is about provenance rather than performance.
How this page is sourced. Every statement about Pindrop below was checked against their own live pages on 31 August 2026 and carries the URL. Not found means checked at that URL on that date and not present — not a claim that they publish it nowhere. Enterprise vendors commonly release validation material under confidentiality during procurement instead. Corrections: corrections@aivoicedetctor.com
Credit where it is due
Four certifications on a live trust centre. SOC 2 Type 2, ISO/IEC 27001:2022, PCI DSS v4.0.0 and HIPAA, at trust.pindrop.com. Truthring holds none of these, and pre-launch is not a mitigating circumstance when a buyer needs an attestation this quarter.
A peer-reviewed submission, with the authors named. A Pindrop paper sits in the ASVspoof 2021 proceedings on the ISCA archive, authored by Tianxiang Chen, Elie Khoury, Kedar Phatak and Ganesh Sivaraman. Submitting to a public challenge means being scored on somebody else’s data by somebody else’s rules, which is a different act from publishing a number about yourself.
Named people, extensively. Chief executive Vijay A. Balasubramaniyan, Dr Paul Judge, Dr Mustaque Ahamad and chief technology officer Neeraj Gupta, alongside the paper authors.
A retention policy with a period in it. Voice features are “retained by Pindrop for a period of up to three (3) years and deleted thereafter”, and “Customer voiceprints are not retained”. Truthring’s periods are not yet set, so on this point they are ahead of us.
A seen versus unseen split. “99% detection of known deepfake engines and more than 90% detection of previously unseen voice cloning models”. Separating generators the model has seen from generators it has not is the most useful breakdown in this field, because the second number is the one that predicts next year. Most vendors publish one blended figure.
Scale, stated. “Over 20 million audio files, including data from 370+ text-to-speech systems”. And a third-party NPR evaluation, which Pindrop cites, reported 81 of 84 audio samples correctly identified — 96.4%.
The criticism, and it is a narrow one
The <1% figure appears without a named test set, without a stated decision threshold and without a base rate. That is the whole of our objection. A false positive rate is not a property of a detector. It is a property of a detector, a test set and a threshold together. Move the threshold and the false positive rate falls while the miss rate rises; change the audio from studio recordings to eight-kilohertz telephony and both move again. Without knowing which audio produced the figure and where the cut was made, two vendors quoting “under one percent” may be describing incomparable things.
The base rate matters just as much and gets discussed even less. On a call floor where perhaps one call in ten thousand is a deepfake, a one percent false positive rate produces roughly a hundred false alarms for every genuine detection. That is not a flaw in the detector; it is arithmetic that applies to every detector including ours. But it is the arithmetic a buyer needs, and a bare percentage does not supply it.
Three further things we looked for that day and did not find: a written methodology document, a stated set of failure conditions, and any accuracy breakdown by recording condition. The 370+ text-to-speech systems are counted rather than named. None of this makes the number wrong; it makes it unverifiable from outside, which is a smaller complaint than the one we make about vendors publishing no error rate at all.
The distinction that actually decides it
Pindrop sells three things and the live versus after-the-fact split runs between them. Pulse scores audio inside a live call — “Detect deepfake audio in just two seconds”. Pulse for Meetings is live and multi-modal, audio and video together, in Teams, Zoom and Webex. Pulse Inspect is the after-the-fact one, run on an uploaded recording.
Truthring only does the last of those, and only on audio. Hand us a live call and there is nothing to hand us; hand us a video and we are analysing one track and blind to the rest of it.
The one place the category runs the other way is access. As at 31 August 2026 we found no self-serve route and no published pricing on pindrop.com; the way in is a sales conversation. For an institution that is normal and often preferable. For one person holding one voicemail at eleven at night, the product simply does not exist.
What each of us publishes
| Published item | Truthring | Pindrop |
|---|---|---|
| False positive rate | Committed, not yet measured. Framework published in advance at accuracy: both error directions, by recording condition. No figure exists yet. | Published. “<1% False positive rate”, footnoted “99.4% accuracy with less than 1% false positives” pindrop.com/product/pindrop-pulse/ · checked 31 August 2026 |
| Test set, threshold and base rate attached to that rate | Committed — dataset described, samples per condition, held-out status stated | Not found alongside the figure at that URL pindrop.com/product/pindrop-pulse/ · checked 31 August 2026 |
| Seen vs unseen generator split | Committed — not yet measured | Published. “99% detection of known deepfake engines and more than 90% detection of previously unseen voice cloning models” pindrop.com · checked 31 August 2026 |
| Accuracy by recording condition Codec, noise, clip length | Committed — the table is published, the cells are unfilled | Not found on the pages checked pindrop.com · checked 31 August 2026 |
| Written methodology document | Published in advance of results, versioned | Not found on the pages checked pindrop.com · checked 31 August 2026 |
| Stated failure conditions | Published | Not found on the pages checked pindrop.com · checked 31 August 2026 |
| Generator coverage | Committed — named and dated | Counted, not named: “over 20 million audio files, including data from 370+ text-to-speech systems” pindrop.com · checked 31 August 2026 |
| Peer-reviewed or challenge submission | None. Pre-launch | Published. ASVspoof 2021 submission; authors Chen, Khoury, Phatak, Sivaraman isca-archive.org/asvspoof_2021/chen21b_asvspoof.html · checked 31 August 2026 |
| Named people | One named publicly — team | Extensively named. Vijay A. Balasubramaniyan (CEO), Dr Paul Judge, Dr Mustaque Ahamad, Neeraj Gupta (CTO) pindrop.com · checked 31 August 2026 |
| Third-party certification | None held. Pre-launch — compliance | SOC 2 Type 2, ISO/IEC 27001:2022, PCI DSS v4.0.0, HIPAA trust.pindrop.com · checked 31 August 2026 |
| Retention period in time units | Not set. Design published, periods unset — data retention | Published with a period. Voice features “up to three (3) years and deleted thereafter”; “Customer voiceprints are not retained” pindrop.com · checked 31 August 2026 |
| Real-time scoring inside a live call | No — after the fact only, by design | Yes. Pulse, “Detect deepfake audio in just two seconds”; Pulse for Meetings is live and multi-modal in Teams, Zoom and Webex pindrop.com · checked 31 August 2026 |
| After the fact, on an uploaded file | Yes — upload a clip | Yes. Pulse Inspect pindrop.com · checked 31 August 2026 |
| Self-serve route and published price | Yes at launch — pricing published in advance | Sales-led. No self-serve sign-up and no published pricing found pindrop.com · checked 31 August 2026 |
Right-hand column checked 31 August 2026 against Pindrop’s own live pages. Not found means checked at the stated URL on that date. Re-checked quarterly.
Where Pindrop is the better choice
- You are stopping fraud, not documenting it. Two seconds inside the call beats an analysis a week later while the money is still moving; at hundreds of thousands of calls a month the job is triage, and that is a platform problem.
- You need signals the audio does not contain. Device, carrier path, account history. Compressed telephony strips exactly the detail an audio-only detector relies on — a weakness we share and they compensate for.
- You need an attested counterparty. SOC 2 Type 2, ISO 27001, PCI DSS and HIPAA in hand today. We have none of them.
- Live meetings are in scope. Pulse for Meetings covers audio and video in Teams, Zoom and Webex. Nothing we build will address that.
- Fifteen years of telephony attack data. Founded 2011, 300+ patents claimed. Exposure at that scale teaches things no public dataset contains.
Where we may be the better choice
- The audio escaped every system. A forwarded voice note, a voicemail, a recording made on a phone. No call flow, no account, no telephony metadata — one file is all there is.
- One person needs one answer today. No procurement, no integration, no minimum volume, no sales call.
- The verdict will be argued about. A method published in advance, an error rate in both directions, a Ringmark reference so the analysis can be pulled up months later, and an explicit unclear when the audio does not support a call.
Every item above is a design commitment made before launch, not a measured result. Nothing on this site claims otherwise.
Questions
Does Pindrop publish a false positive rate?
Yes. Their Pindrop Pulse product page carries “less than 1% False positive rate”, footnoted as “99.4% accuracy with less than 1% false positives”, checked 31 August 2026. That is the disclosure this site argues for, and they make it. What we could not find alongside it was a named test set, a decision threshold or a stated base rate, which is what would let a reader work out what the number means for their own audio. The criticism is about the provenance of the figure, not the figure.
Should a bank contact centre use Truthring instead?
Almost certainly not. A contact centre needs scoring inside the live call, early enough to route it differently, plus signals the audio does not contain — device, carrier path, account history — and case management so a flag becomes an investigation. Pindrop Pulse is built for exactly that and we are not. We analyse one recording after the event.
Which one is more accurate?
We decline to answer, because we cannot. Truthring has not launched and has no measured figures yet; anything we published as a comparison would be a guess. Even between two shipping products the question would need the same held-out clips at the same thresholds, scored by someone with nothing riding on it.
Can an individual buy Pindrop?
Not from the website. As at 31 August 2026 we found no self-serve sign-up and no published pricing on pindrop.com — the route to the product is a sales conversation. If you are one person with one voicemail, that is a practical exclusion rather than a criticism; they sell to institutions.
What does Pindrop publish that Truthring does not have?
A great deal, and it is worth being blunt about it. SOC 2 Type 2, ISO/IEC 27001:2022, PCI DSS v4.0.0 and HIPAA on their trust centre; a peer-reviewed ASVspoof 2021 submission with named authors; extensive named leadership; a retention policy with an actual period in it; and measured accuracy figures of any kind. Truthring holds none of those certifications, has published no measured number, and has not set its retention periods. We are pre-launch and the gap is real.
Sources
- pindrop.com/product/pindrop-pulse/ — false positive rate, detection speed. Checked 31 August 2026.
- trust.pindrop.com — SOC 2 Type 2, ISO/IEC 27001:2022, PCI DSS v4.0.0, HIPAA. Checked 31 August 2026.
- isca-archive.org — ASVspoof 2021 submission — named authors. Checked 31 August 2026.
- pindrop.com — product split, training scale, seen/unseen figures, leadership, retention, absence of self-serve pricing. Checked 31 August 2026.
Claims about Pindrop checked 31 August 2026 against publicly available material. Re-checked quarterly. Corrections are published with a date rather than made silently.