Skip to content

Releases: anand34577/sabhalekh

Release list

Sabhalekh 1.2.0

Sabhalekh 1.2.0 Latest
Latest

Choose a tag to compare

View all tags
@anand34577 anand34577 released this 24 Oct 16:58
v1.2.0
4a57e72
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

First public release.

What Sabhalekh does

  • Records a Teams, Zoom or Meet call (your microphone and the other participants, as two tracks) or imports audio and
    video files, and turns them into a transcript with speaker names and timestamps.
  • English, Hindi and Hinglish in the same meeting. Hindi can be shown as spoken, in Devanagari, or in Roman letters.
  • Speech recognition with NVIDIA Nemotron 3.5 ASR on any graphics card (through Vulkan), on built-in graphics or on the
    CPU. Speaker detection with NVIDIA Nemotron-3-Diarization, up to 8 speakers.
  • Click any word to play from there; rename speakers everywhere at once; edit, merge, split and undo lines.
  • Search across all meetings, including romanised Hindi (kya finds क्या).
  • Export to Word, PDF report, text, Markdown, SRT, VTT, JSON and CSV, one meeting or many at once.
  • Fully offline: no account, no cloud, no telemetry, no update checks. Network access is blocked in code.

Changed since the pre-release builds

  • The app is now called Sabhalekh. The pre-release builds were called MeetScribe, a name already used by other
    products. Your meetings, settings and recordings move to the new data folder (%APPDATA%\Sabhalekh)
    automatically the first time you open Sabhalekh, and playback keeps working for every old meeting.
  • The installer is about a third smaller (about 850 MB instead of 1.3 GB). It no longer includes NVIDIA's cuBLAS
    library, which only the optional Whisper model packs used. Whisper packs still use an NVIDIA graphics card if the
    CUDA 12 Toolkit is installed, or if you copy the two cuBLAS files into %APPDATA%\Sabhalekh\cuda; otherwise they run
    on the CPU. The built-in Nemotron speech recognition is not affected.
  • The hardware badge says where speech recognition runs ("GPU Mode: GTX 1650") without technical precision labels.
  • Settings → About has links to the user guide, the issue tracker and the source code. They open in your browser.

Fixed

  • The welcome guide said "Everything will run on the CPU" on PCs where the graphics card does the speech recognition
    through Vulkan (any PC without NVIDIA's CUDA libraries). It now names the device that is actually used.
  • AMD and Intel graphics cards showed as "No NVIDIA GPU" in the welcome guide and the sidebar.
  • The microphone and speaker test in the welcome guide left a temporary file behind, which then appeared as an
    "unfinished recording" to recover.
  • Talk-time percentages could add up to 101 % (for example 75 % + 26 %). They now always add up to 100 %, in the app
    and in every export.
  • A missing space in the audio test instructions.

Download

File What it is
Sabhalekh-Setup-1.2.0.exe The installer (about 850 MB, speech models included). No admin rights or internet needed.
Sabhalekh-ModelPack-Accuracy.zip Optional: Whisper large-v3 for Hindi-to-English translation. Unzip, then Settings → Models → Import Offline Model Pack.
SHA256SUMS.txt Checksums to verify the downloads.

Runs on Windows 10 and 11 (64-bit). The installer is not code-signed yet, so Windows SmartScreen may ask you to confirm: click More info → Run anyway.

User guide · Report a problem

Assets 5
Loading