Within

Private meeting transcription with timestamps. Your audio never leaves your machine.

Features · Quick Start · Configuration · Contributing · License

Why Within?

Most transcription tools upload your recordings to someone else's server. Within runs OpenAI Whisper locally, so meeting audio stays on your desktop — with optional AI recaps only when you ask for them.

Features

• 🔒 Local-first — runs Whisper on your machine; audio never leaves it.
• 🎙️ Any audio — upload MP3, WAV, M4A, FLAC, OGG, or WebM and track progress in-app.
• ⏱️ Timestamped transcripts — structured segments with quick copy actions.
• 📝 Optional AI summary — bring an OpenAI key to generate bullet notes (only transcript text is sent, never audio).
• 🎛️ Model picker — choose any Whisper checkpoint from tiny to large-v3 to balance speed and accuracy.

More screenshots

Upload & progress

AI summary

How It Works

┌────────────────┐      audio       ┌─────────────────┐
│  React + Vite  │ ───────────────► │  Flask backend  │
│   (frontend)   │ ◄─────────────── │ Whisper (local) │
└────────────────┘    transcript    └────────┬────────┘
                                             │ optional, text only
                                             ▼
                                    ┌─────────────────┐
                                    │   OpenAI API    │
                                    │   (summaries)   │
                                    └─────────────────┘

Transcription happens entirely on your machine. The only outbound request is the optional summary call, which sends transcript text — never audio.

Requirements

• Python 3.10+
• Node.js 18+
• ffmpeg installed and on your PATH (see below)
• First run needs internet to download Whisper weights

Install ffmpeg (required)

Whisper depends on ffmpeg for audio decoding. Install it once:

macOS (Homebrew):

brew install ffmpeg

Ubuntu/Debian:

sudo apt update
sudo apt install ffmpeg

Windows (PowerShell):

winget install Gyan.FFmpeg

Verify:

ffmpeg -version

Quick Start

Clone the repo:

git clone https://github.com/aliffrhn/within.git
cd within

Terminal 1 — backend (Flask + Whisper):

python3 -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt

# Optional overrides
export WHISPER_MODEL=medium   # default: medium
export WHISPER_LANGUAGE=auto  # default: auto
export PORT=5050              # default: 5050

python app.py

Terminal 2 — frontend (Vite):

npm install
npm run dev

Open http://localhost:5173.

The dev server proxies /config, /transcribe, and /summarize to http://localhost:5050. If your backend runs elsewhere, set:

export VITE_BACKEND_URL=http://localhost:5050

Environment Variables (backend)

Variable	Purpose	Default
PORT	Flask port	5050
WHISPER_MODEL	Default checkpoint	medium or recommended
WHISPER_MODELS	Comma-separated allowlist	tiny,base,small,medium,large-v2,large-v3
WHISPER_RECOMMENDED_MODEL	Highlighted model in UI	medium
WHISPER_LANGUAGE	Default language hint	auto
MAX_UPLOAD_MB	Upload limit in MB	40
OPENAI_API_KEY	Server-side summary key	unset
OPENAI_SUMMARY_MODEL	Summary model	gpt-4.1-mini
OPENAI_API_BASE	Custom OpenAI-compatible endpoint	https://api.openai.com/v1

Troubleshooting

• No such file or directory: 'ffmpeg' — ffmpeg is missing or not on your PATH. Install it using the steps above and restart your terminal.
• First transcription is slow — Whisper downloads model weights on first use. Subsequent runs are much faster.
• Out of memory with large models — try a smaller checkpoint like WHISPER_MODEL=small.

Privacy

• Audio never leaves your machine — Whisper runs locally.
• Summary requests send only transcript text plus a language hint.
• Temp files are deleted after transcription completes.

Contributing

Contributions are welcome! Whether it's a bug report, a feature idea, or a pull request — every bit helps.

1. Read the Contributing Guide for setup and workflow.
2. Check open issues or open a new one to discuss your idea.
3. Fork, branch, and send a PR.

Please also see our Code of Conduct.

License

Distributed under the MIT License.

Acknowledgments

• OpenAI Whisper — the speech recognition model that powers Within.
• shadcn/ui & Radix UI — UI components.
• Vite, React, Tailwind CSS, and Flask.

---

If Within is useful to you, consider giving it a ⭐ — it helps others find the project.