Nemotron 3.5 Asr Streaming 0.6b Onnx

--- license: other license_name: openmdw-1.1 license_link: https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b/blob/main/LICENSE.OpenMDW-1.1 base_model: nvidia/nemotron-3.5-asr-streaming-0.6b base_model_relation: quantized pipeline_tag: automatic-speech-recognition library_name: onnxruntime language: - multilingual tags: - onnx - onnxruntime - asr - automatic-speech-recognition - speech-recognition - streaming - fastconformer - rnnt - nemotron --- # Nemotron 3.5 ASR Streaming 0.6B — ONNX ONNX Runtime bundle for NVIDIA Nemotron 3.5 ASR Streaming Multilingual 0.6B. ## Overview ONNX export of [nvidia/nemotron-3.5-asr-streaming-0.6b](https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b), a streaming automatic speech recognition model based on FastConformer + RNNT. It is not a newly trained model. The ONNX artifacts are sourced from [altunenes/parakeet-rs](https://huggingface.co/altunenes/parakeet-rs/tree/main/nemotron-3.5-asr-streaming-0.6b-onnx) and repackaged here for stable, checksum-verified distribution. ## Source | Field | Value | |---|---| | Upstream model | [nvidia/nemotron-3.5-asr-streaming-0.6b](https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b) | | ONNX artifact source | [altunenes/parakeet-rs (nemotron-3.5-asr-streaming-0.6b-onnx/)](https://huggingface.co/altunenes/parakeet-rs/tree/main/nemotron-3.5-asr-streaming-0.6b-onnx) | | Upstream source revision | `24b151a851dd15909e1fc611b11bb2da52b9fc81` | | ONNX artifact source revision | `a95331a1f347c66d68bc7e34d3eb05963bbb2f4c` | | Packaging source revision | `b3ea33d792e4edd1ea9ffe222d250bc3239ee4ae` | | Export tool/script | ONNX redistribution via altunenes/parakeet-rs (checksum-verified repackage) | | Quantization recipe | FP32 ONNX FastConformer + RNNT graphs | ## Files | File | Description | |---|---| | `encoder.onnx` + `encoder.onnx.data` | Cache-aware FastConformer encoder | | `decoder_joint.onnx` | RNNT decoder and joint network | | `tokenizer.model` | SentencePiece tokenizer | | `config.json` | ONNX/export configuration metadata | | `LICENSE.OpenMDW-1.1` | NVIDIA Open Model License | | `NOTICE.md` | Notices | ## Intended Use Streaming/chunk-based multilingual ASR for local inference via ONNX Runtime. The streaming-capable design allows long-form audio to be processed incrementally rather than loading the full audio into memory. This redistribution removes the Python, NeMo, Docker, CUDA Toolkit, and Rust runtime requirements of the original release; it can be loaded directly via ONNX Runtime. ## Runtime Notes - Designed for ONNX Runtime compatible runtimes. - Input: 16 kHz mono audio. - Architecture: FastConformer encoder + RNNT decoder (streaming-capable). - Inference: cache-aware chunked streaming; full-sequence mode also supported. ## Precision and Packaging Export tooling, precision, and quantization are recorded in the **Source** table above. This packaging mirror does not publish independent parity benchmarks; validate on your target execution provider before production use. ## Limitations - Multilingual model; accuracy varies by language and domain. - No repository-specific ASR quality benchmark is documented here. - This is a redistribution of a third-party ONNX export; validate checksums before production deployment. ## License The upstream NVIDIA Nemotron 3.5 ASR model is released under the **NVIDIA Open Model Data License (OpenMDW-1.1)**, not Apache 2.0. See `LICENSE.OpenMDW-1.1` and `NOTICE.md` in this repository. The ONNX artifacts are redistributed in accordance with OpenMDW-1.1. Review the license terms before commercial or production use.