by Nodical

ASDIC is a 2027-2028 research project. Published measurements are those of the notebook; targets are targets.

Learn

Understanding a phone conversation, one page at a time.

What is known, what exists, what we measure. Every page cites its sources and keeps measured facts apart from targets.

  • 2 October 2026

    Why speech recognition fails on the phone

    Narrow band, noise, accents, overlaps, proper nouns, numbers: the six causes of transcription errors on the phone, what each one costs in a ten-second call, and what ASDIC will measure.

    Read →

  • 2 October 2026

    Sovereign AI: what it means for a phone conversation

    Four concrete questions — where the audio travels, who can listen, how long it stays, which law applies — what the GDPR, the CNIL and the AI Act (applicable 2 August 2026) say, and what Nodical already holds, or not yet.

    Read →

  • 2 October 2026

    Pipeline or speech-to-speech: two ways to put an AI on the phone

    A chain of three models (speech recognition, language model, voice) or a single model that goes from speech to speech: what each gains and loses on the phone, what Kyutai and OpenAI publish, and where ASDIC fits.

    Read →

  • 2 October 2026

    Understanding what the caller wants, without going through text

    Spoken language understanding explained plainly: going through text or straight from sound to meaning, what MEDIA, SLURP and LeBenchmark show, their limits, and why an intent in ten seconds is not an intent in three minutes.

    Read →

  • 2 October 2026

    The voicemail beep: why it never comes when you expect it

    Carrier greetings, personal messages, missing or unrecognisable beeps: why a voicemail beep never lands at the same moment, what Twilio's DetectMessageEnd option does, and what it means for leaving a message.

    Read →

  • 2 October 2026

    Ten seconds: what happens when a call is picked up

    The September 2026 measurement of Nodical's calls, told in order: the pick-up, “Allô?”, the human-or-voicemail question, the agent's answer, the hang-up. Only what was measured, and what the measurement does not say.

    Read →

  • 2 October 2026

    Silence on the phone: noise, breath, hesitation

    What voice activity detection (VAD) is, what fools it on the phone, how to read Picovoice's benchmark (Cobra, Silero, WebRTC at 0 dB), and why detecting speech still does not tell you when to answer.

    Read →

  • 2 October 2026

    Phone voice AI in twenty words

    AMD, VAD, end of turn, barge-in, WER, TTFAB, p95, 8 kHz, G.711, full-duplex, SLU, pseudonymisation… The twenty words that come up whenever voice agents on the phone are discussed, each defined plainly and linked to the page that goes further.

    Read →

  • 2 October 2026

    Evaluating a voice AI vendor: ten questions to ask

    Ten neutral questions before choosing a voice agent: which audio, which language, 8 kHz or wideband, median and p95, who measured, where the audio goes, how long it is kept, which subprocessors, how people object, what happens in an outage.

    Read →

  • 2 October 2026

    Spanish on the phone: accents, speed and “¿diga?”

    How people pick up the phone in Spain, Mexico or elsewhere in the Americas, why accents and speed make a machine's job harder, what public leaderboards evaluate in Spanish (and what they do not: the telephone), and why ASDIC has no Spanish figures yet.

    Read →

  • 2 October 2026

    Reading a speech recognition benchmark without being fooled

    Word error rate (WER) worked out by hand, what text normalisation changes, read versus conversational speech, 16 kHz versus 8 kHz, vendor versus third party, average versus total: how to read a speech-recognition figure.

    Read →

  • 1 October 2026

    Detecting voicemail in under two seconds: what exists today

    How an AI recognises an answering machine on the phone: the three methods, what each costs in seconds, what Twilio and research publish, and where ASDIC stands (8 s measured on 60,287 calls, 2 s target).

    Read →

  • 1 October 2026

    End of turn: waiting for silence is not enough

    How an AI knows you have finished speaking: the difference between “no more sound” and “done”, the cut-off / delay trade-off, what LiveKit, Pipecat and Kyutai publish — and what they do not (language, telephone).

    Read →

  • 1 October 2026

    Measuring a voice agent: the five numbers that matter

    A voice agent's “latency” says almost nothing about an outbound call. The five measurements ASDIC publishes (human or voicemail, end of turn, cut-offs, handovers, duration), how to read a median and a p90, and what public benchmarks are worth.

    Read →

  • 1 October 2026

    8 kHz audio: what the models can't hear

    Why speech recognition collapses over the phone: 300–3,400 Hz band, G.711 codecs, lost packets; third-party figures (Whisper, Deepgram on CallHome); and what ASDIC will measure on French calls.

    Read →

  • 1 October 2026

    A French phone-call corpus: why none exists

    An honest inventory of what exists (CALLFRIEND 1996, ESLO, Claire, MEDIA, Luqia's benchmark), why none of it can train an agent that makes calls, and what ASDIC is building and will publish, within the framework set by the CNIL.

    Read →