Real-time AI tool

Copiloto de Reuniones

Local copilot for meetings in English: live transcription, Spanish translation and suggested replies in an always-on-top subtitle overlay, all running on the user's own machine.

Working prototype · v0.1

Request a walkthrough

Project overview

Context and approach

Overview

Copiloto de Reuniones listens to the computer's audio during Meet, Teams or Zoom calls, transcribes the English conversation, translates it to Spanish and proposes a short reply when someone asks a question.

Problem

Following a fast meeting in a second language is hard, and answering a direct question on the spot is harder. Cloud meeting assistants record everything and send it to third parties, which is not acceptable for private or work conversations.

Solution

A local pipeline segments system audio, transcribes it with faster-whisper, translates offline with Argos, detects questions without calling the LLM on every phrase and only then asks a local Ollama model for a short, safe reply. A lightweight client can delegate the heavy work to a GPU server on the same network.

Screenshots

Subtitle overlay floating over the Copiloto main window — The subtitle overlay floating over the main window during the demo meeting.

Subtitle overlay with a detected question and a suggested reply — Always-on-top overlay: detected question, translation and a suggested reply.

Copiloto main window with live transcription and suggested reply — Main window during the built-in demo meeting.

Key figures

347

Test cases

Modes: local or client/server

Raw audio files stored

100%

Local processing

Architecture

01
System audio
Loopback capture of Meet, Teams or Zoom.
02
Segmenter
Voice activity detection; silence is never sent to Whisper.
03
faster-whisper
Streaming English transcription with interim text.
04
Translate + detect
Argos EN to ES offline and a question detector.
05
Suggested reply
Local Ollama model with recent context, shown in the overlay.

Main capabilities

Live English transcription with interim captions while people speak.
Offline English to Spanish translation.
Question detection and short suggested replies with their meaning in Spanish.
Subtitle overlay that floats over the meeting, with a global hotkey and tray icon.
Meeting history, summary and open items when the meeting ends.
Client/server mode over the local network with setup wizard and auto-discovery.

Verification evidence

347 pytest cases across transcription, diarization, memory and UI logic.
Built-in demo mode reproduces a full meeting without a real call.
Packaged as a Windows app with autostart for client and server.

Engineering decisions

— Audio is processed in memory and never written to disk.
— The LLM is only called when a question is detected, keeping latency and load low.
— If the model does not answer, a safe generic reply is suggested instead of nothing.
— Heavy work can move to a GPU server so a modest laptop can still be the client.

Scope limitations

— Prototype focused on meetings held in English.
— Source code is private; screenshots come from the built-in demo meeting.
— Speaker identification by name is still experimental.

Want to see it running?

The repository is private. I can walk you through the code and a live demo.