Live Call Transcription That Drafts a 16-Field Emergency Form While the Operator Listens

#The starting position

Emergency medical consultation lines exist for a simple, important reason. Many people who call an ambulance do not need one, and some who need one hesitate to call. A consultation line lets a member of the public describe a situation to a trained operator, who assesses it and advises whether to call an ambulance, go to a hospital or manage at home. It protects ambulances for the people who need them and gives callers expert reassurance.

The contact centre operator in this engagement runs such a line for a regional public sector client in Japan. Each call is recorded in a structured consultation form with sixteen fields, covering who is affected, what is happening, where, since when and what advice was given.

That form was the problem. Operators had to listen carefully to a worried, sometimes panicked caller, ask the right questions, decide on advice and fill in sixteen fields, all at once. Typing competes with listening. Every moment spent finding a field or correcting a typo is a moment taken from the caller, and on a line like this, attention is the most important thing an operator has.

The operator wanted to move the typing to AI without moving any of the judgment.

#What we built

01 Phase 1: recorded calls and text

The first production system works on recorded audio and text. A recording or transcript is uploaded, transcribed in Japanese, and structured information is extracted and exported to Excel for the client's reporting. This phase proved that transcription and extraction worked on real consultation calls before anything ran live.

02 Phase 2: live call capture

The second production system works during the call. It captures the audio live and produces a running Japanese transcript as the conversation happens.

03 Speaker separation

The system separates the operator's voice from the caller's. On a medical call this is essential: a symptom the caller describes and a question the operator asks must never be confused when the record is drafted.

04 The sixteen-field form, drafted during the call

An extraction step reads the transcript and fills the sixteen fields of the consultation form as the information is spoken. By the time the caller has described the situation, much of the form is already drafted.

05 Operator review and end-of-call summaries

At the end of the call, the operator reviews the drafted form, corrects anything the AI got wrong and saves it. The system also writes an end-of-call summary. The operator remains responsible for every record.

06 Regional vocabulary engineering

After deployment, the biggest remaining source of errors was local vocabulary: place names, landmarks and regional terms that general speech models mishear. We added an administrator-managed vocabulary for each region, deterministic corrections that apply only when the surrounding context makes them safe, an audit of recognition results, and automated regression tests so that fixing one term cannot quietly break another. The latest of these enhancements has been verified in our environment and is being validated on the client's servers before it is treated as live.

#A call, step by step

A caller explains that an elderly relative has had a fall and is now confused. As they speak, the transcript appears for the operator, with the caller's words and the operator's questions clearly separated. The draft form begins to fill: the person affected, their age, what happened, when, the symptoms described and where they are.

The caller mentions a district name and a nearby landmark. The regional vocabulary makes sure both are recognized correctly, because a wrong location on an emergency record is unacceptable. The operator asks follow-up questions, gives advice and stays focused on the caller throughout.

When the call ends, the operator checks the draft. One field needs a small correction. They fix it, save the record and move on, with a summary attached. The structured record is complete, accurate and approved by the person who took the call.

#Why human review is the design, not a safeguard

It would be technically possible to save AI-drafted forms automatically. On an emergency medical line, it would be the wrong choice. Speech recognition will occasionally mishear, extraction will occasionally misplace a detail, and an incorrect emergency record can have serious consequences.

So the system is designed around the operator's review. The AI's job is to draft fast and accurately enough that reviewing is quicker than typing. The operator's job is to make sure the record is right. That division keeps the benefits of AI, an operator who can listen instead of type, while keeping accountability where it belongs.

#Why local vocabulary deserves engineering

General speech models are trained on broad data. They handle common words well and struggle with the specific: district names, small landmarks, local abbreviations and terms that rarely appear outside one region. On a consultation line, those are exactly the words that matter most, because they describe where help may be needed.

Treating vocabulary as an engineering problem made the difference. Managed vocabularies let administrators add terms without code changes. Context-gated corrections apply a fix only when the surrounding words confirm it is appropriate, avoiding new errors. The recognition audit shows which terms still cause trouble. Regression tests run whenever rules change, so accuracy only moves in one direction.

#Introducing AI to operators who did not ask for it

Operators on an emergency line are experienced professionals with a well-practised routine. A new system that changes the screen in front of them during a call has to earn its place, and the design took that seriously.

The live system shows the operator a transcript and a draft form, but it never interrupts or prompts them, and nothing is saved without their action. If the draft is wrong, correcting it takes less time than typing it from scratch would have. If the AI misses something, the operator's own notes and judgment still decide the record.

Introducing the recorded-call phase first helped as well. Operators and supervisors saw the transcription and extraction working on real calls before it appeared during their own live conversations, which made the live system familiar instead of surprising.

#How we delivered it, end to end

Recorded before live. Phase 1 established quality on recordings before Phase 2 touched live calls.

Review built in. Operator review of every record was part of the specification from the start.

Speaker separation as a requirement. The system was designed to keep operator and caller apart in the transcript and the extraction.

Quality engineering after go-live. Real recognition errors seen in production drove the vocabulary, correction and testing work.

Honest release status. Enhancements are verified in our environment and validated on the client's servers before they are called live.

#The decisions that mattered

Assist the operator, never replace them. This is what made AI acceptable on an emergency line.

Phase the risk. Recorded calls first, live calls second.

Engineer the vocabulary. Managed vocabularies, context-gated corrections and regression tests turned recurring errors into solved problems.

#What changed

#What we would tell a public service team considering this

Start with the operator's attention and protect it. Let AI take the typing, keep a trained person approving every record, and introduce the system on recordings before live calls. Then invest in the details that general AI gets wrong in your region, because on an emergency line those details are the ones that matter.