Back to Gap Map Model components / Speech & audio

Nemotron ASR

open weights / Overall score: 4.0(strong)

NVIDIA's streaming speech recognition models, branded under Nemotron rather than Parakeet. Nemotron 3.5 ASR is a 600M-parameter cache-aware FastConformer-RNNT model that transcribes 40 language-locales from one checkpoint, with punctuation, capitalization and configurable chunk sizes down to 80ms; Nemotron ASR Streaming is its English-only predecessor.

The multilingual model is released under OpenMDW-1.1 and the English one under the NVIDIA Open Model License, which sit in the same license tier. Trained and served with NeMo. NVIDIA's batch-first recognizers are parakeet and canary; the Nemotron language models are nemotron.

Openness

3 medium confidence

3.0

weights: open(.nemo and safetensors checkpoints, ungated)
data: partial(Granary, MLS, Common Voice, FLEURS and VoxPopuli are public
code: partial(NeMo inference and streaming scripts and a fine-tuning guide
license: OpenMDW-1.1(Nemotron 3.5 ASR weights)+NVIDIA-Open-Model-License(English Nemotron ASR Streaming weights)

Nemotron ASR's weights download freely under licenses that ask for attribution and impose no limit on who may use them or at what scale. NVIDIA names the public speech corpora in the mixture, but part of it is the proprietary Riva training set, and the card points to NeMo for inference and fine-tuning rather than publishing the recipe that trained this model, so it cannot be reproduced from what is released.

https://huggingface.co/api/models/nvidia/nemotron-3.5-asr-streaming-0.6b recorded 2026-10-27
"gated":false; cardData "license":"other", "license_name":"openmdw-1.1"; files include nemotron-3.5-asr-streaming-0.6b.nemo and model.safetensors.
https://huggingface.co/api/models/nvidia/nemotron-speech-streaming-en-0.6b recorded 2026-10-27
"gated":false; cardData "license":"other", "license_name":"nvidia-open-model-license"; files include nemotron-speech-streaming-en-0.6b.nemo and model.safetensors.
https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b/raw/main/README.md recorded 2026-10-27
"The training data is a dynamic blend of public and proprietary internal datasets ... including: NVIDIA Riva multilingual ASR training set (Proprietary), NVIDIA Granary, Multilingual LibriSpeech (MLS), Mozilla Common Voice, FLEURS, VoxPopuli / Europarl-ASR"; "To train, fine-tune or perform inference with this model, install NVIDIA NeMo"; links the cache-aware streaming inference script and a fine-tuning blog post.

Adoption

4 medium confidence

4.0

Hugging Face downloads summed over the two declared checkpoints, led by the multilingual model. Downloads count file fetches, not transcriptions.

https://huggingface.co/api/models/nvidia/nemotron-3.5-asr-streaming-0.6b recorded 2026-10-27
"downloads":977899 for nvidia/nemotron-3.5-asr-streaming-0.6b.
https://huggingface.co/api/models/nvidia/nemotron-speech-streaming-en-0.6b recorded 2026-10-27
"downloads":107522 for nvidia/nemotron-speech-streaming-en-0.6b.

Capability

not assessed

Nemotron ASR reports word error rates on FLEURS under its own text normalization, while the category places recognizers by the Open ASR Leaderboard, which does not list it. With no shared benchmark against parakeet or qwen-asr, it is left unplaced rather than placed by guesswork.

https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b/raw/main/README.md recorded 2026-10-27
Per-language FLEURS WER tables across chunk sizes, "WER/CER are computed after text normalization that aligns the reference and hypothesis"; no Open ASR Leaderboard result is cited.

Verified 2026-10-27, except some scores, still awaiting re-confirmation