NVIDIA Nemotron Speech Streaming EN 0.6B — ONNX FP16 export (80ms latency)
=======================================================================

Licensed by NVIDIA Corporation under the NVIDIA Open Model License

This repository is an ONNX (FP16) export of the original NeMo checkpoint produced
directly from NVIDIA's official Hugging Face repository. It is a derivative work
distributed under the same license that governs the source model.

Source model
------------
  Repository : nvidia/nemotron-speech-streaming-en-0.6b
  URL        : https://huggingface.co/nvidia/nemotron-speech-streaming-en-0.6b
  Revision   : 7a9b763e6c5fb103da690219c049fac917aa50b1

Conversion details
------------------
  Export method   : NeMo toolkit ASRModel.export() to ONNX
  ONNX opset      : 17
  att_context_size: [70, 0]  (80ms streaming latency)
  Precision       : FP16 weights with FP32-typed graph I/O (Cast-wrapped).
                    The cache-aware pre_encode subsampling front-end is retained
                    in FP32 for converter compatibility; the 24-layer Conformer
                    body runs in FP16.
  Chunk geometry  : chunk_mel_frames = 8, pre_encode_cache_frames = 9,
                    total_input_frames = 17, chunk_audio_samples = 1280

License
-------
  NVIDIA Open Model License Agreement (see LICENSE.txt for the full text).
  URL: https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/

  NOTE: This model is governed by the NVIDIA Open Model License.
  It is NOT licensed under CC-BY-4.0 (a tag that has appeared on some
  third-party exports of this checkpoint and is incorrect).

Toolkit versions used (full list in conversion/dep_baseline_nemo_export_env_macos.txt)
--------------------------------------------------------------------------------------
  python 3.12.13
  nemo_toolkit 2.7.3
  torch 2.12.1+cpu
  torchaudio 2.11.0+cpu
  onnx 1.22.0
  onnxruntime 1.27.0
  onnxconverter-common 1.16.0
  sentencepiece 0.2.1
  numpy 1.26.4

Reproducibility
---------------
  The export and conversion scripts (export_nemotron.py, convert_fp16.py) are
  included in the conversion/ folder of this repository.