Offline Model Packs

Offline model packs

The installer already contains everything you need: NVIDIA Nemotron 3.5 ASR for speech recognition and Nemotron-3-Diarization for telling speakers apart. Model packs are optional extras you can add without an internet connection on the PC that runs Sabhalekh.

The Accuracy pack

Adds	OpenAI Whisper large-v3
Gives you	Translate Hindi to English, and custom vocabulary support
Size	about 1.5 GB
Download	Sabhalekh-ModelPack-Accuracy.zip on the releases page

You only need it for translation. For normal English, Hindi and Hinglish transcripts, the built-in Nemotron model is more accurate on Hindi and several times faster.

Installing a pack

1. Download Sabhalekh-ModelPack-Accuracy.zip. You can download it on any computer and carry it over on a USB drive.
2. Right-click the zip → Extract All… You get a folder called Sabhalekh-ModelPack-Accuracy.
3. In Sabhalekh, open Settings → Models and click Import Offline Model Pack…
4. Choose the extracted folder (not the zip) and click Select Folder.
5. Sabhalekh checks every file against its checksum and copies it into %APPDATA%\Sabhalekh\models. That takes a minute or two. When it's done, Whisper large-v3 shows as installed.

You can delete the zip and the extracted folder afterwards.

To use translation, set Settings → Transcription → Language to Translate Hindi to English, or pick it for one meeting with ⋯ → Re-process…

Making Whisper fast on an NVIDIA graphics card (optional)

Whisper is a large model. On the CPU, translating an hour-long meeting can take a long time. On an NVIDIA graphics card it's many times faster, but that needs a library from NVIDIA called cuBLAS, which Sabhalekh can't include in its installer for licensing reasons.

You can check whether it's set up in Settings → Hardware → NVIDIA CUDA (Whisper packs). If it says Available, you're done. If not, pick one of these:

Option A: install the CUDA Toolkit (simplest). Download the CUDA Toolkit 12.x for Windows from NVIDIA's website and install it with the default options. Restart Sabhalekh. (Version 12 is needed; 13 won't work.)

Option B: copy just the two files you need (smaller). If you already have them from another program, or download NVIDIA's cuBLAS 12 redistributable, copy cublas64_12.dll and cublasLt64_12.dll into this folder (create it if it doesn't exist):

%APPDATA%\Sabhalekh\cuda

Restart Sabhalekh, then click Settings → Hardware → Re-detect.

The built-in Nemotron model never needs cuBLAS. It uses your graphics card through Vulkan, which comes with every modern graphics driver.

Making your own pack

Developers can bundle their own models, for example a fine-tuned Nemotron or another Whisper size, with tools\make_manifest.py. See Building from source.