orangeblue39/nemotron-3.5-streaming-en-cuda

SpeakEasy Nemotron 3.5 Streaming English — CUDA

This repository contains the SpeakEasy-derived float32 ONNX pack used by the Windows CUDA streaming-ASR path. It is an application-specific export, not an official NVIDIA ONNX release and is not endorsed by NVIDIA.

The upstream checkpoint is NVIDIA's nemotron-3.5-asr-streaming-0.6b, pinned to revision f3d333391852ba876df169dcc9ba902d25b6ab0b. SpeakEasy exported the 320 ms float32 ONNX files with the pinned environment in scripts/model-export using NeMo revision 23de54bb6d5c1c4abd97c8d51b8c20e91447dd32.

Downloaded pack

The application downloads this archive and verifies its length and SHA-256 before extraction or activation:

File	Bytes	SHA-256
nemotron-3.5-streaming-en-cuda-320ms-fp32.tar.gz	2406690010	5f6a28fe4d33b038f5062dc064bc55260d35278e64c5f600d4908b5c524f1717

The archive contains:

• encoder.onnx — 42247579 bytes — 288271e199371ea7fbd0ba7e246f243d4e7685a7670618194698c19e030967a1
• encoder.data — 2454405120 bytes — 7584f85df76bc9ae6fbdfa53aa8d97b07a842525d1c501d536d77fd9e4f57ac7
• decoder.onnx — 59764943 bytes — ab1934b40bffacfc4fd6795fb1b5c6531cd4403cbf434e0144f69fc8aba3f8c2
• joiner.onnx — 37824290 bytes — 02dd5450d1ca9c541ceb037b94b09a9b262b4a4c22c240a97537f59ac250d338
• tokens.txt — 144528 bytes — 32be3ebfabfff475d64d7829b435f1c7856a1c497907def5c41d54ca9f1eccfd

The pack targets sherpa-onnx 1.13.4, the CUDA execution provider, Windows x86-64, 16 kHz English transcription, and true-online transducer decoding.

License and notices

The model materials are provided under OpenMDW-1.1. The associated sherpa-onnx runtime is Apache-2.0. See NOTICE.md for attribution and derivation details.

This export is distributed without user audio or application source code.