AI Translator

Every word, bothways, in seconds.

Live interpretation for the consultation itself. Both sides transcribed as they speak, carried across in either direction, and the clinical note written in one language at the end of it — with no interpreter to book and nobody's relative translating in the corridor.

A chest-pain history, mid-consultation

Consultation · chest pain · room 4

Interpreting
DoctorListeningENEnglish
PatientHablandoESEspañol

You

When did the chest pain start?

MédicoInterpreted

¿Cuándo empezó el dolor en el pecho?

PatientInterpreted

Last night, after dinner. It hurts more when I breathe deeply.

Usted

Anoche, después de cenar. Me duele más cuando respiro hondo.

Pleuritic character recorded · red-flag set armed

You

Does it spread to your arm or your jaw?

MédicoInterpreted

¿Se le extiende al brazo o a la mandíbula?

PatientInterpreted

No, only here. But I get short of breath going up stairs.

Usted

No, solo aquí. Pero me falta el aire al subir escaleras.

Exertional dyspnoea + pleuritic pain → same-day review per protocol

Session

EN ⇄ ES · auto-detected

Median latency

420 ms

Interpreter booked

Not required

Disposition

Same-day review

One consultation, two microphones, each side in its own language
languages and dialects, detected from how someone speaks
60+languages and dialects, detected from how someone speaks
from the end of a sentence to the other side hearing it
~400msfrom the end of a sentence to the other side hearing it
note, in one language, however many were spoken
Onenote, in one language, however many were spoken

How it runs

Nobody picks a language from a list

The moment a consultation needs an interpreter is the moment it stops being a consultation and becomes a logistics problem. This removes the logistics rather than speeding them up.

  1. It works out the language

    From the first few words, not from a menu and not from what the record says the patient speaks. A record that is three years out of date is one of the more common reasons an interpreter is booked for the wrong language.

  2. It carries both directions at once

    The clinician speaks and the patient hears it in theirs; the patient answers and the clinician hears it in English. Both halves are transcribed as they go, so the consultation has a record rather than a memory.

  3. The note comes out in one language

    Whatever was spoken, the clinical note is written in the language your record is kept in, and the patient's instructions go out in the language they were speaking. Nobody translates anything twice.

Why clinical is different

General translation is not safe here

A consultation is not a document. It switches language mid-sentence, it turns on a word like “radiates”, and the hedge a patient puts around a symptom is frequently the most clinically useful thing they say.

  • It follows a switch mid-sentence

    Bilingual patients move between languages inside one sentence, especially under stress and especially for symptoms. The model runs token by token rather than sentence by sentence, so it goes where the speaker goes instead of waiting for a full stop.

    • Token-level detection
    • No menu, no reset
    • Handles code-switching
  • It knows the clinical register

    Anatomy, medication names, dose units, red-flag phrasing and the difference between describing and reporting. A general translation engine will render "it radiates" as something a triage protocol cannot act on.

    • Anatomy and dosing
    • Red-flag phrasing
    • Units preserved
  • It does not summarise the patient

    Interpretation is turn for turn. Hedges, uncertainty and the bits that do not fit the question are carried across rather than tidied away — because what a patient almost says is often the thing the clinician needed.

    • Turn for turn
    • Hedges preserved
    • Nothing condensed
  • Both sides are on the record

    The original and the delivered line are both kept against the encounter, so a question about what was actually said has an answer. That is the part a family member interpreting in the corridor can never provide.

    • Original retained
    • Translation retained
    • Auditable per turn

Coverage

60+ languages and dialects. These are the ones that carry the most traffic.

  • ES
  • PL
  • DE
  • FR
  • ZH
  • AR
  • IT
  • PT
  • UK
  • HI
  • and 50+ more

When it should stop

An interpreter that knows when it is not enough

The risk in machine interpretation is not a wrong word. It is a wrong word delivered with the same confidence as a right one. Everything here is built around making that visible in the moment rather than in a complaint six weeks later.

  • Consent and the fact of machine interpretation are stated to the patient at the start
  • Low confidence on a clinically material turn stops and asks rather than guessing
  • A human interpreter can be brought onto the same call without restarting it

The limits, up front

Three things it will not do

  • It does not replace a certified interpreter where one is required

    Consent for a procedure, a safeguarding conversation, a breaking-bad-news appointment: where your policy or your jurisdiction requires a certified human interpreter, the room says so and helps you get one rather than proceeding.

  • It does not guess when it is unsure

    A turn that carries clinical weight and comes back below the confidence threshold is flagged in the moment — repeated back for confirmation rather than passed on as though it were certain.

  • It does not interpret the clinical meaning

    It translates what was said, not what it thinks was meant. Deciding that chest pain radiating to a jaw is a cardiac presentation is the clinician's job, and the transcript exists so that judgement is made on real words.

FAQ

Before you ask.

How is this different from a phone interpreter line?

Latency and availability, mostly. A three-way call with an agency interpreter takes minutes to convene, turns a ten-minute consultation into thirty, and is often unavailable for the less common languages at the hours you need it. This is in the room from the first sentence, in any of 60+ languages, with no booking. Where policy requires a certified human, it helps you get one instead of replacing them.

What happens if it mishears something important?

Confidence is scored per turn. A clinically material turn below threshold is not passed on as fact — it is flagged and repeated back for confirmation, in both languages, before the consultation moves on. Every original and every delivered line is retained against the encounter, so a later question about what was said is answerable rather than a matter of recollection.

Which language does the clinical note end up in?

The one your record is kept in — English by default, configurable per tenancy. The patient's instructions and any follow-up message go out in the language they were actually speaking. Neither is a translation of the other after the fact; both are generated from the same transcript.

Does it handle dialects and code-switching?

Yes, and code-switching is the normal case rather than the edge case. Bilingual patients routinely describe symptoms in their first language inside an otherwise English sentence. Detection runs token by token, so it follows the switch rather than waiting for the sentence to end and getting the whole thing wrong.

Where is the audio processed, and is it kept?

In EU infrastructure under your tenancy, and not by default. Interpretation works from the live stream; what persists is the transcript and the note, on the retention schedule you set. Nothing is used to train third-party foundation models, and identifiers are scrubbed in memory before anything is written.

Bring us the language you cannot cover

The one where the interpreter is a two-day wait, or the one three families in your catchment speak and nobody on the rota does. We will run a real consultation in it while you listen.