Back to Portfolio
TALENT EXHIBITION - PROMA2
WhispFlow Icon

WhispFlow

I built WhispFlow as a desktop app that records meetings, turns speech into searchable notes, and creates structured summaries. It runs locally by default, and you can plug in cloud or self-hosted engines for transcription and AI. It was my PROMA2 Talent Exhibition project.

Plain-English view, built for everyone Full technical detail
2
ways to transcribe: local engine or an API
8
AI providers, from cloud to fully local
Separate
setups for live captions and full transcripts
Local
by default: audio leaves only if you choose an API

How WhispFlow works

Record-to-notes pipeline
1 Capture mic · system · hybrid · file record live, or import an audio file
2 Transcribe local engine · API on your machine, or through Groq, OpenAI or your own server, chosen per run
3 Summarize 8 providers · cloud or local summaries, action items and cleanup, from the transcript text only
4 Export PDF · DOCX · TXT take the transcript, summary and action items with you
Live captions local or API set up separately from full transcripts, shown while you record

Everything runs locally by default. Audio leaves your machine only if you pick a transcription API, and AI features only ever send the transcript text, never the audio.

What WhispFlow does

Records any meeting or lecture

One click captures your microphone, the system sound, or both, and you can also import an audio file.

Turns speech into text, your way

On your own computer with no internet needed, or through a transcription service when you want more speed. You choose for each recording.

Summaries with action items

AI pulls out the key points, decisions and to-dos so you leave with clear notes instead of a long recording.

Private by default

Everything stays on your machine unless you pick an online service. Live captions and full transcripts are set up separately, so you decide what goes where.

About This Project

I built WhispFlow for the Talent Exhibition (PROMA2 course) in the BTS Cloud Computing programme at Lycée Guillaume Kroll. I chose a real documentation problem, designed a working solution, and applied the cloud and AI concepts from the semester.

The Problem

During meetings and lectures, important information often gets lost. Taking notes manually splits attention between listening and writing, and reviewing long recordings afterward is time-consuming. I wanted a simple way to leave each session with useful notes instead of another recording to search through.

The Solution

I built WhispFlow to turn a meeting or lecture into organized notes: record once, get a transcript, then receive key points, decisions, and action items ready to share. Everything runs on the computer by default, and you choose whether transcription or summaries use an online service.

Under the Hood

Desktop App

ElectronNode.jsMediaRecorderHTML/CSS/JS
  • Desktop shell with audio capture via the MediaRecorder API, session management UI, live waveform visualization, and organized file storage.
  • Session management with search, tags, sorting, and context menus

Transcription Engines

faster-whisperwhisper.cppGroq / OpenAIself-hosted
  • A built-in local engine (faster-whisper or whisper.cpp, models from tiny to large) or any OpenAI-compatible transcription API: Groq, OpenAI, or a self-hosted server. Chosen per run, with a fallback to the local engine if an API fails.
  • Local or API transcription, chosen per run

AI Summarization

OpenAI-compatibleOllamaLM Studiocached
  • One AI gateway for summaries in four styles, action items and transcript cleanup. Works with OpenAI, Mistral, Anthropic, Groq, Cohere, Ollama, LM Studio or any OpenAI-compatible endpoint, with model discovery. Only the transcript text is sent.
  • AI summaries from eight providers, cloud or local

Local Storage & Versioning

FilesystemJSONWebM/Opusversioned
  • Session folders hold the audio, versioned transcripts, and cached summaries. Everything stays on your device in an organized structure that's easy to back up and export.
  • Re-transcribe with a different model or language - no data is overwritten. Switch between transcript versions (v1, v2, v3...) via dropdown. Each version tracks the model, language, and timestamp. Summaries are keyed by transcript version and language (e.g. v1_en, v2_fr).

Core Workflow

01

Record - One-click recording or keyboard shortcut (Ctrl+R). Three modes: Microphone, System Audio, or Hybrid. Audio files can also be imported. Live waveform and timer during recording.

02

Transcribe - Pick the engine for each run: the local engine (faster-whisper or whisper.cpp, tiny to large models) or a transcription API such as Groq, OpenAI or a self-hosted server. Live captions have their own setting, so they can use a different engine than full transcripts.

03

Summarize - AI generates key points, decisions, action items and a full summary in four styles. Choose from eight providers, cloud or fully local, or add any OpenAI-compatible endpoint. Generate in any language.

04

Export - Export to PDF, DOCX or TXT with the transcript, summary and action items. Share with colleagues or classmates.

Lessons Learned

Hybrid architecture works

Local by default for privacy and offline use, with cloud engines as an option, so each user picks their own balance of speed, cost and privacy.

Model selection matters

Letting users choose between speed and accuracy (tiny vs. large Whisper models) solved the hardware-limitation problem elegantly.

Never overwrite user data

Versioning means you can always go back. Users trust the tool more when they know nothing gets deleted.

Cache everything expensive

Caching AI summaries by transcript version and language prevents needless API calls and keeps costs under control.

Project Phases

Phase 1 (Done) - Local-First: Desktop recording, local transcription, AI summaries, versioning, session management and export.

Phase 2 (In progress) - Open Engines: Transcription APIs (Groq, OpenAI, self-hosted), separate live captions, a local fallback when an API fails, and one provider list shared by transcription and AI are working. Cloud storage, sync, accounts and collaboration are still planned.