← AICédric MerlinFREN

AI galaxy Case 3 of 8

An assistant that prompts me in meetings

The problem

During a meeting, there is no time to look up a figure or an answer.

The constraint

The conversation must not go to the cloud.

What I did

01

Everything is heard, with consent

The microphone and the computer’s audio are captured continuously, with the participants’ consent.

02

Transcribed on the machine

Whisper transcribes the conversation on the computer, sending nothing outside.

03

Jargon corrected

A corrector restores the right technical terms (Wazuh, Cloudflare, Scaleway).

04

Three lines, at the right time

The AI whispers three lines to me, only when I need them.

Diagram: from sound to three whispered lines, all on the machineA microphone and the computer’s audio enter a computer as waves, where they become words; a crossed-out cloud shows that nothing leaves. A badly transcribed word, Wazu, is corrected to Wazuh. Three lines slide into the margin, with a delay of 5 to 8 seconds.miccomputeraudioWazuWazuhnothing leaves5 to 8 s

The result

Transcription error rate down from 5.65% to 0.43%. 16 business terms out of 16 recognised, on my test bench. 5 to 8 seconds of delay.

5.65 %, then 0.43 %

transcription error rate

Stack

  • whisper.cpp
  • Swift (ScreenCaptureKit)
  • ffmpeg
  • LLM