by NodicalASDIC is a 2027-2028 research project. Published measurements are those of the notebook; targets are targets.
NotebookEnd of October 2026: the measurement, on the person's voice alone
From 28 to 30 October 2026, on 6,642 real calls, pick-up detection hung up on 62.7% of voicemails in 2.66 seconds (median): 86.5% in Spanish, 59.3% in French.
What changes in the method
Until now, the reference label (person, voicemail, silence) was read from the full recording, where the agent's voice mixes with the person's: the labelling sometimes confused the two. From this measurement on, the label is read from the person's voice alone, recorded separately during the first twenty seconds. These figures are therefore not directly comparable with the previous entry.
The call set
6,642 answered calls, drawn at random from those of 28, 29 and 30 October 2026, on two call servers, in French and Spanish: a first batch of 4,688 calls, then a control batch of 2,000 other calls used for nothing else. 27 October is excluded: a line incident made the start of calls silent that day, which would have skewed the measurement.
The method
For each call: the decision actually taken by the detector (engine log), a word-by-word transcription of the person's voice alone, then automatic labelling by a language model. No human labelling at this stage.
Results in production
| Measurement | Overall | Spanish | French |
|---|---|---|---|
| Calls | 6,642 | 1,220 | 5,422 |
| Voicemails hung up on at pick-up | 62.7% | 86.5% | 59.3% |
| Time to that decision (median / 9 in 10) | 2.66 s / 3.43 s | 2.41 s / 3.44 s | 2.79 s / 3.43 s |
| People taken for a voicemail, at most | 3.4% | 1.1% | 5.2% |
“At most”: reading the French cases one by one, many are in fact personal greetings or call screening that the automatic labelling took for people. The true rate is lower; it is not measured yet.
What it tells us
Spanish is well covered. In French, four voicemails in ten still get through: some say nothing during the first two seconds, others start with a lone “Bonjour.”, and many greetings start with the name recorded by the subscriber, a pause, then the operator's voice (“is not available at the moment”): the detector took that name followed by silence for a person.
The test bench
To fix this without risking hanging up on a person, we built a bench that replays the real decision rule on these recordings, without making a single call. Three changes were tried there and put into service on 31 October: a lone “Bonjour.” no longer proves a person is speaking; a lone name followed by a pause waits for the rest of the text; eleven greeting phrasings were added, each checked against 995 labelled people and 16,536 conversations with people, without touching any. On the bench, these changes recognise about 6 to 7 more points of voicemails, on the first batch and on the control batch alike, with no extra person lost. This is a replay, not a production measurement.
What does not work yet
- Labelling is still automatic: a sample labelled by ear is needed to know the true share of people taken for a voicemail.
- The bench underestimates production (the recorded track stops at the moment of the real decision): it compares two rules, it does not give a production figure.
- The target is still a target: under 2 seconds, with under 3% error. It is not reached.
Next
Measure in production, with the same method, the effect of the rules put into service on 31 October; build a sample labelled by ear; publish both here.
And in practice? Nodical campaigns
Corrections
No correction so far.