Sabhalekh

Download for Windows · Website · User guide · Report an issue

Sabhalekh (सभालेख, "meeting record") records your Teams, Zoom or Google Meet calls, or takes an audio or video file, and gives you a transcript that says who spoke, when, and what they said. It understands English, Hindi and the mix of both that most of us actually speak.

Everything happens on your own computer. There is no account, no cloud service and no upload. Your recordings never leave the PC.

Why Sabhalekh

• Private by design. Speech recognition, speaker detection and search all run locally. The app blocks its own network access in code, and the tests fail if anything tries to go online.
• Made for Indian meetings. English, Hindi and Hinglish in one transcript, with Hindi in Devanagari or in Roman letters, your choice. Search finds क्या when you type kya.
• Knows who is talking. Up to 8 speakers per meeting. Your own microphone is labelled "You", and you can rename everyone else with one click.
• Runs on ordinary laptops. On a 4 GB GTX 1650 a one-hour meeting is ready in about 5 minutes. Any graphics card works (NVIDIA, AMD or Intel), and so does a plain CPU.
• Free and open source. MIT licensed. No ads, no subscription, no telemetry.

Features

Record or import

• Records your microphone and the other participants as two separate tracks
• Works with Teams, Zoom, Meet in the browser, or anything else that plays sound
• Global hotkey (Ctrl+Shift+R) and a tray indicator while recording
• Crash-safe: an interrupted recording is recovered on the next start
• Import mp3, wav, m4a, aac, flac, ogg, wma, mp4, mkv, mov, avi or webm by drag and drop

Read and fix the transcript

• Click any word to play the audio from there; the current word is highlighted
• Rename a speaker everywhere at once, or move a single line to another speaker
• Edit, merge and split lines; undo any edit
• Talk time per speaker
• Find inside a meeting (Ctrl+F) or across all meetings from the sidebar

Export anywhere

• Word (.docx), PDF meeting report, plain text, Markdown
• Subtitles (.srt, .vtt) for the recording
• JSON with word timings, CSV for spreadsheets
• With or without timestamps and names; many meetings into one ZIP

Sensible about hardware

• Picks the fastest device by itself: graphics card, built-in graphics, or CPU
• If the graphics card runs out of memory, it moves to the next device and carries on
• Frees all memory after each meeting, so it doesn't slow down your calls
• Optional Whisper model pack for Hindi to English translation

Install

1. Download Sabhalekh-Setup-x.y.z.exe from the latest release.
2. Run it. It installs for your user account only, so it needs no admin rights, and it needs no internet connection.
3. Open Sabhalekh. A short welcome guide checks your microphone and speakers and records a 10-second test.

The installer is large (about 850 MB) because the speech models are inside it. After that, nothing else is ever downloaded.

Windows SmartScreen may warn you because the installer is not code-signed yet. Click More info → Run anyway. Every release lists SHA-256 checksums in SHA256SUMS.txt, so you can check that your download is the published file.

Requirements: Windows 10 or 11 (64-bit), 8 GB RAM or more, and about 2 GB of free disk space. A graphics card makes it faster but isn't required. Windows 10 needs the Microsoft Edge WebView2 Runtime, which most PCs already have.

Upgrading from a pre-release build called MeetScribe? Install Sabhalekh, open it once, and your meetings move over automatically. Then uninstall MeetScribe from Windows Settings → Apps.

How to use it

To…	Do this
Record a meeting	Click New Recording (or press Ctrl+Shift+R from any app), join your call, then Stop & save
Transcribe a file	Click Upload Media, or drag files onto the window
Name the speakers	Click "Speaker 1" in the transcript and type the name
Jump in the audio	Click any word or timestamp
Share the result	Export → pick a format

Please tell people when you record a meeting. In many places it's required by law, and it's always polite.

The user guide explains every step with pictures, from installing to troubleshooting.

How it works

Part	What it uses
Speech recognition	NVIDIA Nemotron 3.5 ASR (0.6B) on transcribe.cpp, on any GPU through Vulkan or on the CPU
Who spoke when	NVIDIA Nemotron-3-Diarization, ported to ONNX + NumPy, on the CPU at about 40× real time
Voice detection	Silero VAD v6
Translation (optional)	OpenAI Whisper large-v3 through CTranslate2, from the Accuracy model pack
App	Tauri 2 + React + TypeScript; Python 3.12 backend (FastAPI, SQLite with full-text search)

The reasoning behind each choice, with measurements, is in docs/design-decisions.md.

Build from source

You need Windows 10/11, Python 3.12, Node.js 20+ and Rust. Then:

git clone https://github.com/anand34577/sabhalekh.git
cd sabhalekh
py -3.12 -m venv .venv
.venv\Scripts\pip install -r backend\requirements-dev.txt
.venv\Scripts\python -m pytest tests -q          # unit tests (no models needed)
cd app; npm ci; npm test; cd ..

Running the app or building the installer also needs the converted models and the transcribe.cpp runtime. The Building from source page walks through it.