Nemotron 3 diarization — Windows ARM64 QNN hybrid v1
This repository contains the fixed-shape, calibrated W8A16 QDQ representation
qualified for @goodnotesglobal/speaker-identification-node on Windows ARM64.
QNN owns the transformer backbone. The 13 nodes under /model/head/* remain
FP32 on CPU because quantizing that head changed speaker-turn decisions.
The publisher verifies the graph and external-data sidecar against the
qualified byte sizes and SHA-256 digests before creating a Hugging Face commit.
manifest.json records the pinned upstream model, calibration provenance,
transformations, tool versions, and output hashes. Device-specific QNN EP
contexts are compiled by the installed app and are not portable assets.
Source model: onnx-community/Nemotron-3-Diarization-ONNX at
353b6f8ad2cac3580e982d7fbdf0a010786b0406.