Case 01 · Allergy centre · healthcare
Call speech analytics for an allergy centre
About 36,000 minutes of calls a month are recognised by our own model, fine-tuned on medical terminology. Conversations are checked against the centre's standards.
The allergy centre handles about 36,000 minutes of calls a month across about ten phones. Nobody can listen to that volume by hand, so the system reviews staff conversations with patients.
Speech is recognised by our own model, fine-tuned on medical terminology: it doesn't confuse the names of drugs, procedures and diagnoses. Each recording is transcribed and scored against the centre's standards. Management gets ready exports and reports per staff member and sees how quality changes over time.
The project in numbers
- About 36,000 minutes of conversations a month.
- Our own speech model, fine-tuned on medical terminology.
- About ten phones from which recordings are collected.
- Management exports are built automatically, without manual preparation.
What the system does
- Collects call recordings from the centre's telephony itself. The telephony has no open API, so we wrote a dedicated parser for the export.
- Builds a card for every call: the conversation transcript, a score, remarks and recommendations.
- Calculates scores by call and by conversation stage and aggregates them per staff member.
- Shows quality trends, strengths and weaknesses in reports, and prepares a daily report per staff member.
- A manager can adjust a score and a comment manually.
- A web interface with filters, reports and recording playback. Data can be passed to other systems.
Where AI works
- Turns a call recording into text. This is done by our own speech model, fine-tuned on medical terminology.
- Analyses the conversation transcript:
- gives an overall score,
- splits the conversation into stages,
- identifies the topic, reason, goal and outcome of the call,
- lists objections and recommendations.
- Finds typical mistakes in unsuccessful calls.
- Flags recurring problems in daily and summary reports.
Planned
- Integration with the clinic's information system: a call is linked to an appointment in the schedule, showing whether the patient actually came.
- Winning back lost patients: a list of those who called but didn't book, with the reason, so a receptionist can call back the same day.
Project interface
Screenshots are taken from the current interface version on demo data, in Russian: a call card, a manager report, a call list.
Fortune AI
Have a similar task?
In a free consultation, we'll discuss your process and possible uses of AI.