by NodicalASDIC is a 2027-2028 research project. Published measurements are those of the notebook; targets are targets.
NotebookOctober 2026: the pick-up detector
On 7,075 answered calls from 14 to 16 October 2026, the detector running in Nodical recognised 64.8% of voicemails at pick-up, in 2.41 seconds (median), and 77.5% over the whole call.
The question
A first ASDIC building block is already running in Nodical's engine: it listens to the start of the call and says whether it is a person or a voicemail. How many voicemails does it recognise, at what moment, and how many people does it wrongly take for a voicemail?
The call set
7,075 answered calls from Nodical campaigns, 14 to 16 October 2026, in French and Spanish. Version measured: the detector as it ran before the 18 October fixes.
The method
For each call: the engine log (moment and type of decision), a new transcription of the call, then automatic labelling by a language model (person, voicemail, uncertain). No human labelling at this stage.
The results
| Measurement | Overall | Spanish | French |
|---|---|---|---|
| Voicemails recognised at pick-up | 64.8% | — | — |
| Time to that decision (median) | 2.41 s | — | — |
| Voicemails recognised over the whole call | 77.5% | 75.5% | 80.3% |
| People taken for a voicemail at pick-up | 6.2% | 5.3% | 12.5% |
Call length when the voicemail is recognised: 5.5 s; when it is missed: 24.9 s. The share of people taken for a voicemail is probably overstated (see below).
What it tells us
The pick-up decision comes much earlier than in September (2.41 s versus 8 s median), but it still misses one voicemail in three at the start. A missed voicemail costs nearly twenty seconds more on average. And some people are taken for a voicemail, especially in French: that is the most serious error, since it cuts a real conversation.
What does not work yet
- Labelling is automatic: it sometimes mistakes the agent's voice for the person's, which inflates the share of people taken for a voicemail. Human labelling of a sample is needed.
- Three days, two languages, our own campaigns: these figures describe Nodical at that date, not “detectors”.
- The target is still a target: under 2 seconds, with under 3% error. It is not reached.
Next
Fixes were deployed on 18 October (distorted voicemail phrases, greetings contradicted by a first human word). Their effect will be measured in production with the same method, then on a hand-labelled sample, before being published here.
And in practice? Nodical campaigns
Corrections
No correction so far.