---
license: other
license_name: mixed-upstream
license_link: https://github.com/strigov/babki
---

# Babki model bundles

Signed, self-contained model bundles for the Babki desktop transcription
application. The application downloads exactly the archives referenced by its
signed release (SHA-256 and size are pinned inside the installer); nothing here
is meant to be used directly.

## Contents

| Path | What | Upstream | License |
|---|---|---|---|
| `bundles/babki-quality-cpu-v*.zip` | Quality (CPU): GigaAM v3 e2e-RNNT ONNX FP32, pyannote Community-1, closed CPU-only Python 3.11 runtime with torch 2.8.0+cpu | see below | see below |
| `bundles/babki-quality-gpu-v*.zip` | Quality (GPU): Community-1 voice model with its CUDA runtime; uses GigaAM and other models from the basic pack | see below | see below |
| `models/gigaam-v3-e2e-rnnt-onnx-fp32-v1.tar` | GigaAM v3 e2e-RNNT ONNX FP32 export | [ai-sage/GigaAM-v3](https://huggingface.co/ai-sage/GigaAM-v3), ONNX conversion [istupakov/gigaam-v3-onnx](https://huggingface.co/istupakov/gigaam-v3-onnx) | MIT |
| `models/pyannote-community-1.tar` | pyannote speaker-diarization-community-1 at revision `3533c8cf8e369892e6b79ff1bf80f7b0286a54ee`, unmodified | [pyannote/speaker-diarization-community-1](https://huggingface.co/pyannote/speaker-diarization-community-1) | CC-BY-4.0 (weights), MIT (pyannote.audio) |

Exact upstream revisions, SHA-256 values and the conversion command are in
the `*.artifact-release-record.json` files next to each model and inside every
bundle under `notices/`. The Python runtime inside the Quality bundles is
assembled from published wheels (PyPI and download.pytorch.org); their exact
filenames and SHA-256 values are listed in
`notices/python-wheelhouse-inventory.json`, and each wheel's own license text
ships in `Lib/site-packages/*.dist-info`.

## Attribution

- GigaAM: Copyright (c) 2024 GigaChat Team. MIT.
- Thanks to Ilya Stupakov for the GigaAM ONNX conversion for onnx-asr.
  Its LICENSE.txt is identical to the upstream GigaAM MIT notice above.
- pyannote speaker-diarization-community-1: pyannote (Hervé Bredin et al.),
  CC-BY-4.0. No copyright notice was supplied. Offered as-is, without
  warranties; see [Section 5 of CC BY 4.0](https://creativecommons.org/licenses/by/4.0/legalcode#s5).
  Naming these projects does not mean that their authors or licensors endorse Babki.
  Segmentation, embedding and PLDA files are used without modification.
  This mirror redistributes the unmodified weights with attribution
  as the license requires; the original gated page collects contact details
  for the pyannote newsletter, which this mirror does not do.
- Silero VAD v6: Silero Team, MIT.
- ONNX Runtime: Microsoft, MIT. FFmpeg (LGPL build): FFmpeg team.
- CPython 3.11 (python-build-standalone): PSF / MPL-2.0. PyTorch: BSD-3-Clause.