# Attribution and third-party notices

## Source model

- Work: `nvidia/parakeet-tdt_ctc-0.6b-ja`
- Creator: NVIDIA NeMo team
- Source: https://huggingface.co/nvidia/parakeet-tdt_ctc-0.6b-ja
- Pinned revision: `44edb27eea9317daf89333e75eb830db4b1cc298`
- Source checkpoint: `parakeet-tdt_ctc-0.6b-ja.nemo`
- Source checkpoint SHA-256: `a2ed2aab9c82cce4e9b699dca669184d7525c7b5e1eb8c3f7b789421bc3637e8`
- License: Creative Commons Attribution 4.0 International
- License URL: https://creativecommons.org/licenses/by/4.0/

The source model is a Japanese hybrid FastConformer TDT-CTC model trained by
NVIDIA using ReazonSpeech v2.0, as described in the upstream model card.

## Modifications

The source model was converted from NeMo to split ONNX graphs. A shared encoder
was extracted from the CTC export, and constant-weight `MatMul` operations in
that encoder were dynamically quantized to signed 8-bit per-channel weights.
The TDT decoder/joint and CTC head remain FP32. Full details are recorded in
`MODIFICATIONS.md` and the finalized `distribution-manifest.json`.

This is an unofficial derivative. NVIDIA has not reviewed, sponsored, or
endorsed this conversion.

## Build tools

The conversion uses NVIDIA NeMo, PyTorch, ONNX, and ONNX Runtime. These build
tools are not redistributed as part of the model payload. Exact versions are
listed in the distribution manifest.