On the phone, everything happens in seconds. Off-the-shelf models were built for read speech, in English. We are building our own, for phone conversation, in French, Spanish and English, trained and hosted in Europe.
In development · research project 2027-2028
“Today” values are measured on our September 2026 calls. Targets are targets: they will be published with the measurements, not before.
Speech recognition models learned from read speech, in wide band, mostly in English. On the phone the sound is narrow, people cut each other off, and a decision must land in under a second. Every second of understanding lost costs a contact.
Read from Nodical's production database, 14 campaigns, without reading the content of conversations.
ASDIC says who the agent is talking to, with a confidence score. Today the decision lands after eight seconds (median). The target: under two seconds.
Interested, not interested, call back later, wrong number, asks for a human, does not want to be contacted again: the intent is recognised live, along with the answers to eligibility questions. Matched against the sales the customer actually closed, these signals refine each contact's chances of winning: the team knows who to focus on.
Neither cutting them off nor leaving a gap. Waiting for silence is not enough: it is the weak point of every voice agent, and the heart of our work.
When an adviser should take over, ASDIC says so, with a two-line summary so nobody arrives cold.
French, Spanish, English: the same model, the same performance, without rebuilding one model per language.
| Challenge | Today (measured) | Target |
|---|---|---|
| Human or voicemail | decision at 8 s (median), 14 s for 9 calls in 10 | under 2 s, under 3% error |
| Response delay | 0.14 s median, 1.32 s 9 times out of 10, excluding end-of-speech detection | under 0.5 s, cutting in less than 5% of the time |
| Handover | 1.19% of answered calls, at 135 s (median) | more than 9 handovers in 10 judged relevant |
| Intent and eligibility | no measurement without an annotated corpus | set after the baseline |
| Chances of winning | computed in Nodical since October 2026 from the call outcome and duration, based on each customer's sales; no published measurement yet | rank better who signs, using signals from the conversation; target set after a one-month pilot |
| Sovereignty | 0% own model | 100% for the understanding model |
Targets proposed from the September 2026 measurements and the state of the art; they will be validated with the team we hire.
ASDIC is the ancestor of sonar, developed during the First World War with the decisive contribution of French physicist Paul Langevin. It sends a sound, listens to the echo and decides within seconds whether it hears a real target or just noise.
Our model does the same thing on the phone: it listens, tells a human from a voicemail at once, spots a real intent in the middle of the conversation, and raises its hand at the right moment.
The model, its weights and its corpus belong to Nodical. It is trained on European computing power and hosted in the European Union. No single vendor can switch it off.
The open reference model for speech recognition gets 2.7% of words wrong in read speech, but 13.8 to 17.6% in telephone conversation.
OpenAI, Whisper paper (table 8)The average time Twilio's answering-machine detector announces with default settings, “close to 100%” accuracy in the United States, lower internationally. No figure published for French or Spanish voicemails.
Twilio, AMD documentation (FAQ)How often you cut someone off by answering after 300 ms of silence. To get down to 5%, you have to wait 1.6 seconds.
LiveKit, eot-bench benchmarkNo public reference measurement exists for French telephone speech. Building that corpus is part of the research.
A corpus of pseudonymised calls, annotated in French, then Spanish and English. None exists in open access for French telephone speech.
Off-the-shelf models measured on this corpus: the quantified starting point.
Design and training of the first four functions.
Spanish and English, accents included.
Integration into Nodical and trials in real conditions, on live campaigns.
Final measurements, published; protection of the model; roll-out.
With the support of a public research laboratory in the Paris region specialised in speech processing, currently being selected. Training on European computing power.
ASDIC listens and understands. It never produces a voice and never learns to imitate one.
It helps the agent lead the conversation, and the customer's team choose the order of its callbacks. It makes no decision imposed on the person called, who never sees its result.
Calls used for training are pseudonymised: names, numbers and addresses are removed before any annotation. Anyone can object by a simple email.
ASDIC replaces the “listen and judge” part of Nodical's voice agents. The voice is still produced by speech synthesis.
Phone reception, healthcare, public services, emergency lines: any system that receives or places calls in French, Spanish or English.
A model hosted in the European Union, with its weights and its corpus, for organisations that cannot send their conversations elsewhere.
The project is led by Douglas Demart, president of Nodical. Two engineers join the team in Paris: the first in January 2027 (French model), the second in 2027 (multilingual and real time). A real corpus of more than 60,000 calls a month, European computing, a partner laboratory, the possibility to publish.
Nodical calls, answers and calls back companies' customers, billed by the second, with no subscription. ASDIC will become its ears and judgement. In the meantime, here is what Nodical does today.
The agent calls, qualifies and hands over only the contacts worth it. ASDIC's measurements are taken on these calls.
24/7 receptionEvery incoming call is answered, qualified and passed on, at any hour, in any language.
What we doCall-back in under 60 seconds, follow-ups, appointments, reports: what Nodical builds for any company with customers to call.
An email is enough. We answer ourselves, and we publish nothing that has not been measured.
by Nodical