Steno

Your private stenographer

Steno is the AI powered intelligence layer for all your confidential conversations, your private data never leaves anywhere. Record, transcribe, summarize, and query your meetings using local AI models. Perfect for government, defence, legal and C-suite professionals with confidential data needs.

_{Trusted by users at AWS, Deliveroo, Tesco, Hashicorp, Rutgers & European Union.}

_{Sponsored by Gitlab Founder's Open Core Ventures.}

_{Disclaimer: This is an independent open-source project for meeting-notes productivity and is not affiliated with, endorsed by, or associated with any similarly named company.}

If you're looking for a hosted desktop recording API, consider checking out Recall.ai, an API that records Zoom, Google Meet, Microsoft Teams, in-person meetings, and more.

• 2026-06-09 🪟 Windows alpha — Steno now runs on Windows 10/11 (x64). Record, live Parakeet transcription, summaries, and system-audio diarisation all work; ships as an unsigned installer for now (SmartScreen warns on first launch), with transcription on CPU via ONNX Runtime.
• 2026-06-09 ☁️ AWS Bedrock as a cloud AI provider — Bring your own Bedrock API key to route summaries through Claude on AWS. Application inference profile ARNs work too, so governed environments with IAM policies that allow bedrock:InvokeModel on application profiles only are supported out of the box.
• 2026-06-09 📅 Calendar polish round — Cancellable Google + Outlook OAuth from the calendar nudge, declined events drop at the IPC boundary (covers organiser-then-declined edge case), 3-per-page < / > pagination on busy days, full-card rendering for all-day blocks, and a 2-minute background poll that survives route changes.
• 2026-06-07 🎙️ Live transcription with Parakeet TDT v3 — Real-time on-screen transcripts during recording via Apple Silicon's MLX backend. Sentences appear as you speak in a Granola-style chat-bubble view; speakers are attributed to You vs Others in real time.

• Privacy-first — 100% on-device; your recordings, transcripts, and summaries never leave your Mac.
• Live transcription with speaker labels — Real-time on-screen text as you speak via Parakeet TDT v3 on Apple Silicon (MLX). Granola-style chat-bubble view with [You] vs [Others] attribution live during the recording and on the final transcript.
• Auto start/stop meetings — Steno notices when a meeting starts and offers to take notes, then offers to summarise when it ends. Granola-style frictionless capture.
• System audio capture — Record both sides of virtual meetings, headphones on, no extra setup or virtual cable. Native Core Audio Tap on macOS 14.4+.
• In-app note-taking — Jot notes while you record; they're folded straight into the AI summary alongside the transcript.
• Ask your meetings — Natural-language Q&A across a single note or your entire library via the Chat tab. Pulls from summary, key topics, and the full transcript.
• Multi-language (25 live, 99 total) — Parakeet covers 25 European languages with live transcription; Whisper handles 99 languages including Chinese, Japanese, Arabic, and Hindi post-stop.
• Markdown ownership — Summaries and transcripts save as clean Markdown you can edit, search, or sync to whatever knowledge base you live in.
• Bring your own cloud model — Optional OpenAI, Anthropic, AWS Bedrock (Claude — including application inference profile ARNs for governed AWS environments), or custom API endpoint for users who prefer a hosted LLM.
• Organisation AI — On managed deployments, sign in to your org's Steno adapter and AI routes through it automatically — no local API key, no per-user setup.

Transcription Models:

• Parakeet TDT v3 (572 MB): Highest quality, supports live transcription, 25 European languages (English, Spanish, French, German, Italian, Portuguese, Dutch, Russian, Polish, Czech, and 15 others). Apple Silicon only via MLX. (default on fresh installs)
• Whisper Large V3 Turbo (1.6 GB): Best accuracy Whisper model. 99 languages including Chinese, Japanese, Arabic, Korean, and Hindi. Post-stop only.
• Whisper Small (466 MB): Balanced speed and accuracy. Same 99-language coverage. (legacy — kept available for existing users; Large V3 Turbo recommended)

Summarization Models (Ollama):

• llama3.2:3b (2GB): Fast and lightweight for quick meetings (default)
• gemma3:4b (2.5GB): Lightweight and efficient
• qwen3.5:9b (6.6GB): Excellent at structured output and action items
• deepseek-r1:14b (9.0GB): Strong reasoning and analysis capabilities
• gpt-oss:20b (14GB): OpenAI open-weight model with reasoning capabilities

• Live transcription during recording
• NVIDIA Parakeet as a transcription engine option
• Editing notes after processing
• Windows version

Download the latest release (Apple Silicon Mac, macOS 12 Monterey or later):

• Apple Silicon (M1-M5)

Intel Mac users: v0.4.0 is Apple Silicon only. Stay on v0.3.8 — the last release supporting Intel Macs. Auto-update on existing Intel installs will not push v0.4.0 to those machines.

1. Download and open the DMG file

2. Drag the app to Applications

3. When you first launch the app, macOS may show a security warning

4. To fix this warning:

   • Go to System Settings > Privacy & Security and click "Open Anyway"

   Alternatively:

   • Right-click Steno in Applications and select "Open"
   • Or run in Terminal: xattr -cr /Applications/Steno.app

5. The app will work normally on subsequent launches

You can run it locally as well (see below) if you don't want to install a DMG.

Windows 10/11 (x64) is supported in alpha, with the full pipeline verified working: record → live Parakeet transcription → batch transcript → Ollama summary, including system-audio loopback capture with [You]/[Others] diarisation.

Install: download stenoAI-windows-x64.exe from the latest release and run it — it installs per-user (no admin needed) and creates Start-menu/desktop shortcuts. On first launch, Windows SmartScreen warns because the alpha is unsigned — click More info → Run anyway. The first-run setup wizard then downloads the transcription model (~670 MB) and the summarisation model (~2 GB).

Before the first tagged Windows release lands, grab the installer from the Windows build workflow: sign in to GitHub, open the latest green run, download the stenoai-windows artifact, and run the .exe inside.

Known alpha limitations:

• Unsigned — SmartScreen warns on first launch; we'll code-sign before 1.0.
• CPU-only summarisation — the bundled Ollama runs on CPU (the NVIDIA GPU libraries are excluded to keep the download small); a separate GPU build is a follow-up. Transcription is CPU on every platform regardless.
• Auto-update is wired (NSIS + latest.yml) but updates are unsigned until code signing is in place.
• Transcription runs through onnx-asr (ONNX Runtime) instead of MLX, with the same Parakeet model and behaviour as macOS. Whisper is also available as an engine option.

Issues + feedback welcome on the GitHub issues tracker.

• Python 3.9+
• Node.js 18+

git clone https://github.com/ruzin/stenoai.git
cd stenoai

# Backend setup
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

# Download bundled binaries (Ollama, ffmpeg)
./scripts/download-ollama.sh

# Build the Python backend
pip install pyinstaller
pyinstaller stenoai.spec --noconfirm

# Frontend
cd app
npm install
npm start

Note: Ollama and ffmpeg are bundled - no system installation needed. The setup wizard in the app will download the required AI models automatically.

cd app
npm run build

stenoai/
├── app/                  # Electron desktop app
├── src/                  # Python backend
├── website/              # Marketing site
├── recordings/           # Audio files
├── transcripts/          # Text output
└── output/              # Summaries

Setup wizard debug console: during first-time setup, expand the debug console panel to see real-time logs of model downloads and service startup.

Terminal logging (recommended for runtime issues): launch the app from a terminal to stream all logs (Python subprocess output, Whisper transcription, Ollama API traffic, error stack traces):

/Applications/Steno.app/Contents/MacOS/Steno

System Console:

# View recent Steno-related logs
log show --last 10m --predicate 'process CONTAINS "Steno" OR eventMessage CONTAINS "ollama"' --info

# Monitor live logs
log stream --predicate 'eventMessage CONTAINS "ollama" OR process CONTAINS "Steno"' --level info

• Update didn't install: Auto-updates are applied on next quit. Quit via the Steno → Quit menu (not just closing the window), then reopen.
• No system audio / no [Others] speaker labels: macOS needs Screen Recording permission. Go to System Settings → Privacy & Security → Screen & System Audio Recording, enable Steno, and relaunch the app.
• stenoai:// deep link doesn't start recording: Make sure Steno has launched at least once after install so the URL scheme is registered. If it still fails, check the terminal log for Protocol handler registration output.
• Recording stops early: Check microphone permissions, Screen Recording permission (if using system audio), and available disk space.
• "Processing failed": Usually an Ollama service or model issue — check the terminal logs.
• Empty transcripts: Whisper couldn't detect speech — verify audio input levels.
• Slow processing: Normal for longer recordings; Ollama is CPU-intensive. If summaries are unusually slow, switch to a smaller model in Settings → AI (Llama 3.2 3B is the fastest).

• User Data: ~/Library/Application Support/stenoai/
• Recordings: ~/Library/Application Support/stenoai/recordings/
• Transcripts: ~/Library/Application Support/stenoai/transcripts/
• Summaries: ~/Library/Application Support/stenoai/output/

This project is licensed under the MIT License.