nemotron-3.5-asr-streaming-0.6b-windows-onnx

Nemotron 3.5 ASR Streaming 0.6B — Windows ONNX

This repository contains Goodnotes-built, fixed-shape ONNX representations of NVIDIA Nemotron 3.5 ASR Streaming Multilingual 0.6B for the direct C++ Windows runtime in @goodnotesglobal/parakeet-windows-node.

The immutable source checkpoint and tool versions are recorded in manifest.json. portable-fp16 serves Windows x64 accelerators and CPU recovery. qnn-w8a16-hybrid quantizes the encoder and joiner for Qualcomm HTP while retaining the fused LSTM decoder in FP32 on CPU. portable-fp32 is an explicit evaluation-only export and is omitted from production publications.

The default streaming window is 560 ms. The pinned checkpoint was trained for 80, 320, 560, and 1,120 ms cache-aware attention contexts; the exporter rejects other sizes. In particular, the separately qualified 2,240 ms Apple model configuration is not a valid Windows ONNX export target for this checkpoint. Compare any supported candidate with 560 ms on identical manifests before promotion. On Windows ARM64, the candidate is W8A16 QDQ (uint8 weights and uint16 activations) for encoder and joiner, with device-specific QNN EP-context compilation performed ahead of recording and persisted locally. Those contexts are hardware/runtime-specific and are not distributed as portable model files. Windows x64 uses the type-validated portable FP16 artifact only after the target provider passes quality, latency, and energy qualification; the same artifact provides CPU recovery.

No Sherpa Node runtime is required. Consumers must validate provider placement, quality, latency, and energy on their target hardware. See the upstream NVIDIA model repository for the governing model license and acceptable-use terms.