wordpipe-nemotron-fast-fp32-projected

Wordpipe Nemotron 3.5 ASR Streaming Fast Profile

This repository contains a Wordpipe-specialized ONNX profile derived from nvidia/nemotron-3.5-asr-streaming-0.6b. NVIDIA is the upstream model developer. Wordpipe adds export, graph specialization, packaging, and local desktop runtime integration; this profile is not a separately trained checkpoint.

This repository publishes the fast Wordpipe profile:

• Build profile: fp32-projected
• Runtime description: FP32 projected-cache model; fastest validated profile, largest footprint.

It is consumed by Wordpipe with:

wordpipe model-install --profile fast --prebuilt-repo fractalyzer/wordpipe-nemotron-fast-fp32-projected

Files

The repository root contains the runtime ONNX profile files:

tokenizer.model
encoder.onnx
encoder.onnx.data        # when the encoder uses external ONNX data
decoder_joint.onnx
decoder_joint.onnx.data  # when the decoder uses external ONNX data
config.json              # when produced by the build pipeline

wordpipe-model-profiles-manifest.json records file sizes, SHA-256 hashes, source directories used when publishing, and the Wordpipe build profile. MODEL_SPEC.md documents the graph specializations, runtime ABI assumptions, and bundled reproducibility scripts used to derive these artifacts.

Intended Use

These artifacts are intended for local Wordpipe desktop dictation on Linux. They are optimized for the Wordpipe Parakeet/Nemotron runtime layout and are not a general NeMo checkpoint replacement.

Evaluation Status

Wordpipe validates candidate profiles with local WER and real-time-factor tests before promotion. Current Wordpipe release validation is primarily English LibriSpeech-based unless a release explicitly says otherwise. See the Wordpipe project documentation and release notes for the exact benchmark set used for a given upload. Do not read the upstream NVIDIA FLEURS numbers as measurements of these transformed artifacts.

Limitations

This profile is packaged for CPU-oriented Wordpipe usage and may not match NeMo's standard runtime interface. The compact profile intentionally publishes the ONNX graph; the Wordpipe installer converts it to ORT format locally for startup-time behavior.

License and Attribution

The upstream model card states that use of nvidia/nemotron-3.5-asr-streaming-0.6b is governed by the OpenMDW 1.1 license. Review the upstream NVIDIA model card and OpenMDW terms before redistribution or deployment. This repository preserves that attribution and publishes derived inference artifacts for Wordpipe.