One frozen research cohort
1,445 historical recordings, with binary human and machine labels. The same recordings were used for the baseline and guarded configuration. Corpus balance is shown above.
PERFORMANCE / RESEARCH EVIDENCE
The evidence
behind the answer.
Voicemail detection accuracy should come with a denominator.
These figures come from one completed research replay in the AMD v2 evaluation environment. Explore what the guarded candidate achieved, how it improved human recognition, and the timing behind the result.
96.68%
Overall accuracy
1397 / 1,445 correct
95.97%
Human recall
119 / 124 human recognized
96.74%
Machine recall
1278 / 1,321 machine recognized
0
Unknown results
0 execution errors in this replay
Historical binary-label research, not a current production accuracy promise. The guarded AMD v2 research candidate is a separate configuration from the frozen KP02 audio-only engine; do not attribute these scores to KP02. Machine-heavy cohort: 124 human and 1,321 machine recordings. No new labeled holdout or production-call qualification is represented here.
IMPROVEMENT IN OUR OWN RESEARCH
More people
recognized as people.
The guarded candidate recognized eight more humans than the AMD v2 baseline in the same historical cohort.
HUMAN RECOGNITION / SAME 124 RECORDINGS
AMD v2 baseline89.52%
111 / 124 humans
Guarded candidate95.97%
119 / 124 humans
+6.45percentage points
in human recall
in human recall
| Metric | AMD v2 baseline | Guarded candidate |
|---|---|---|
| Overall accuracy | 96.33% | 96.68% |
| Human recall | 89.52% | 95.97% |
| Machine recall | 96.97% | 96.74% |
| Human → machine | 13 | 5 |
| Machine → human | 40 | 43 |
| Median audio decision | 3.22 s | 10.32 s |
| P95 audio decision | 4.74 s | 12.94 s |
Guarded candidate: +0.35 percentage points overall accuracy, fewer human-to-machine errors, three more machine-to-human errors, and longer decision time. Both are Kooya research configurations.
TIMING / THE AUDIO CLOCK
A decision needs
evidence.
Audio-prefix time is how much opening audio had become available before a judgment. It is separate from server compute time, network time and webhook delivery time.
The report used accelerated database replay. These values are not live end-to-end or provider wall-clock latency.
GUARDED CANDIDATE / MEDIAN10.32s
50th percentile audio decision point
GUARDED CANDIDATE / P9512.94s
See timing fields in the result →95th percentile audio decision point
THE COMPLETE PICTURE
Every outcome
has a place.
Overall accuracy counts both error directions. Balanced accuracy gives human and machine recall equal weight: 96.36% on this cohort.
| Actual label | Predicted human | Predicted machine | Unknown |
|---|---|---|---|
| Human · 124 | 119correct | 5human → machine | 0 |
| Machine · 1,321 | 43machine → human | 1278correct | 0 |
Read accuracy in context.
1,321 of 1,445 recordings were machine-labeled. Strong overall accuracy can hide a weaker human result, so both recalls and error counts stay visible.
Historical labels are not a fresh three-class adjudicated holdout. This report does not establish call-screener, language-specific, carrier-specific or codec-specific performance.
Unknown remains a valid service outcome even though this replay recorded none. Applications should define a safe fallback for it.
METHOD / PROVENANCE
A result you can trace.
Completed causal replay
Accelerated replay consumed available audio and stored evidence. The run completed its expected result set. No freshly dialed production calls are represented by this page.
All terminal outcomes count
Accuracy is correct / all cases. Human and machine recalls use their own class denominators. Unknowns and execution errors stay explicit rather than being removed.
Prove it on your workflow
Agree on representative audio, labels, language/codec slices, error tolerances and timing goals. Validate the selected deployed service configuration before live call routing.
- Report date
- 2026-09-16
- Aggregate extraction checked
- 2026-10-08
- Run ID
f4263296-796f-4ab7-b000-d78d93b6da3d- Source report SHA-256
3ccfc097886950a6e17d148727a71ca4545a51ea262b0df5c4ad42a7d995ae66