A percentage travels badly
It is measured on one kind of audio and then repeated everywhere, without the conditions that produced it. By the time it reaches a slide, it means nothing.
Measurements & limits
What we measured.
What we will not say.
This page holds the things VoxShield can show you — speed, capacity, cost and test discipline — and the list of things it deliberately does not claim.
There is no accuracy figure on this page, on purpose. A single percentage hides the audio, the language and the conditions it came from, and it is the first thing a buyer misreads. Detection test runs, their datasets and their caveats live in the project repository, where they can be read with their conditions attached — ask us for them.
Speed
How long one check takes
Measured 2026-09-21 on one laptop — RTX 4050 (6 GB), CUDA 12.4, one process — over 20 runs after 3 warm-ups. Each check covers a 4-second slice of audio.
| Step | Runs on | Measured | Budget |
|---|---|---|---|
| Synthetic-voice check | GPU | 16.7 ms | 180 ms |
| Same check, no GPU | CPU (4 threads) | 279.1 ms | — |
| Speaker check | GPU | 4.9 ms | 25 ms |
| Memory used, both models loaded | GPU | 85 MB of 6 081 MB | ≈ 582 MB |
| Whole pipeline, end to end | — | 308 ms | — |
The honest phrasing is “a first verdict in under five seconds, refreshed about once a second”. We avoid the words “real-time”, because the system works on slices of audio, not instantly.
Capacity
64 calls at once, nothing dropped
Measured 2026-09-22. The rule was set before the test: the largest number of calls where 95% of results still arrived within 1.5 seconds and nothing was dropped.
| Calls at once | Verdicts produced | 95% arrived within | Dropped |
|---|---|---|---|
| 1 | 26 | 1.057 s | 0 |
| 8 | 200 | 1.112 s | 0 |
| 24 | 624 | 1.127 s | 0 |
| 48 | 1 248 | 1.210 s | 0 |
| 64 | 1 664 | 1.168 s | 0 |
Call 65 was refused on purpose, with a clear error. The supported wording is “at least 64 — that is where we stopped testing, not where it breaks”. We never say “handles N calls”.
The build
Checks on the code itself
351Tests passing1 skipped.
91.97%Code covered by testsAgainst an 80% floor.
15 / 15Model files verifiedChecked at startup; a changed file stops the app.
₹0Cost21 dependencies, all permissive licences.
Exactly one part of the code reaches the internet, and only when you install it — the step that downloads the models. A test checks that this stays true.
The missing number
Why there is no accuracy percentage here
High scores can be worthless
In our own testing, a candidate that flagged nearly every clip as fake produced a spectacular recall number — while being useless, because it flagged genuine speakers too.
So the product refuses instead
Rather than publish a number we would have to defend in every condition, the system answers only where it has been checked and stays silent everywhere else.
The full test runs — what was tested, on which datasets, with which caveats — are kept in the project repository rather than advertised here, and can be shared with anyone evaluating the work.
Limits
What VoxShield does not claim
This list is not an apology. It is what makes the rest of the page worth reading.
| Item | Status | Detail |
|---|---|---|
| Accuracy figure | Not claimed | No percentage is advertised, in any language or condition. |
| Language coverage | Not claimed | The system answers only inside a narrow validated case and refuses elsewhere. |
| Replay attacks | Not claimed | Detecting a recording played back down the line is not implemented. |
| Live carrier integration | Not claimed | Built against the published protocol and tested against a local mock only. |
| Bank integration | Not claimed | Nothing here has ever held or blocked a real payment. |
| How signals are combined | Rule of thumb | The weights were chosen by judgement, not fitted to data. Calibrating them is future work. |
| Capacity ceiling | Measured with a caveat | 64 is where testing stopped, not where the system fails. The load came from one machine. |
| Two of our own signals | Built, switched off | Built, tested, and left out of the live result because they did not earn their place. |
| Certification | None | No ISO, SOC 2 or equivalent, and no third-party audit. |
| Incidents, automatic response, app hand-off | Designed only | Specified in detail, not built. See the current-stage table on the home page. |
House rule
Never say / say instead
The rule the team holds itself to — including on this website.
| Never | Instead |
|---|---|
| “99% accurate” | Nothing — we do not publish an accuracy figure |
| “Works in every language” | “It answers where we have checked, and refuses elsewhere” |
| “Handles N calls at once” | “We measured 64 on this laptop, this way” |
| “Real-time” | “First verdict under 5 seconds, refreshed about once a second” |
| “Blockchain-secured” | “A hash-chained log that shows if it was altered” |
| “Integrated with banks” | “Not built, and never tested against a real one” |
| “Detects replay attacks” | “We do not claim replay detection” |
Evidence
Everything here is reproducible
Speed, capacity and test numbers regenerate from the repository on the laptop named above. The detection test runs are available on request, with their conditions attached.