nemotron-3.5-asr-0.6b-onnx

nemotron-3.5-asr-0.6b-onnx

ONNX export of nvidia/nemotron-3.5-asr-streaming-0.6b for onnx-asr, model type nemo-conformer-rnnt.

The source model is a cache-aware streaming FastConformer encoder with an RNN-T decoder and language-ID prompt conditioning. This export runs the model in offline (full-utterance) mode: the whole utterance goes through the encoder in one pass with the chunked-limited attention mask of the largest lookahead (att_context_size = [56, 13], 1120 ms chunk). Streaming caches are not part of this graph. Transcripts are identical to NeMo/Transformers offline inference with the same lookahead — see Parity.

The language prompt is frozen to auto (automatic language detection, prompt id 101), so the graph takes audio only and needs no runtime changes. The model appends a language tag such as <en-US> after the final punctuation. Strip it if you only want the words:

import re
text = re.sub(r"\s*<[a-z]{2,3}-[A-Z]{2}>\s*$", "", text)

Usage

import onnx_asr

model = onnx_asr.load_model("nemo-conformer-rnnt", "OpenVoiceOS/nemotron-3.5-asr-0.6b-onnx")
print(model.recognize("audio.wav"))

# int8
model = onnx_asr.load_model("nemo-conformer-rnnt", "OpenVoiceOS/nemotron-3.5-asr-0.6b-onnx", quantization="int8")

Requires onnx-asr with the nemo128_raw preprocessor. This checkpoint uses NeMo normalize: NA — raw log-mel features, no per-feature normalization. The stock nemo128 preprocessor normalizes and produces degraded transcripts. config.json sets "normalize": false so a supporting runtime picks the raw preprocessor automatically. Support lives in TigreGotico/onnx-asr branch feat/nemotron.

Files

Parity

4 FLEURS en_us test clips, CPU, compared with transformers offline inference of the source checkpoint at the same lookahead (13). fp32 ONNX output is character-identical on all 4 clips (apart from the language tag, which the Transformers tokenizer strips as a special token). int8 is close but not identical — a few word and punctuation errors — so prefer fp32 when accuracy matters.

Languages

40 language-locales, same as the source model: 19 transcription-ready, 13 broad-coverage, 8 adaptation-ready. See the source model card.

Limitations

• Offline mode only. Streaming (chunked encoder with cache tensors) is not exported.
• The language prompt is frozen to auto. A different language needs a new export.
• Accuracy in offline mode is the accuracy of the 1120 ms chunk setting, not of an unlimited-context offline model — the encoder always uses limited left/right attention context.

License

Governing Terms: OpenMDW-1.1, inherited from the source model. The full text is in LICENSE.

The source model is Copyright NVIDIA Corporation. This repository redistributes converted weights; the original model card, and its bias, explainability, privacy and safety statements, are at nvidia/nemotron-3.5-asr-streaming-0.6b.