nemotron-3.5-asr-arabic-dialectal-v2-onnx

ONNX export of oddadmix/nemotron-3.5-asr-arabic-dialectal-v2 for onnx-asr, model type nemo-conformer-rnnt.

The source model is a dialectal-Arabic fine-tune of nvidia/nemotron-3.5-asr-streaming-0.6b — 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. The language prompt is frozen to ar (prompt id 7), so the graph takes audio only, needs no runtime changes, and emits no language tag.

Usage

import onnx_asr

model = onnx_asr.load_model("nemo-conformer-rnnt", "OpenVoiceOS/nemotron-3.5-asr-arabic-dialectal-v2-onnx")
print(model.recognize("audio.wav"))

# int8
model = onnx_asr.load_model(
    "nemo-conformer-rnnt", "OpenVoiceOS/nemotron-3.5-asr-arabic-dialectal-v2-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

File	Size
encoder-model.onnx + encoder-model.onnx_data	2.4 GB
decoder_joint-model.onnx	95 MB
encoder-model.int8.onnx	617 MB
decoder_joint-model.int8.onnx	24 MB

Parity

5 FLEURS ar_eg 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 5 clips. int8 is close but not identical — a few word errors and one truncated utterance — so prefer fp32 when accuracy matters.

Accuracy of the source fine-tune (WER 0.423 on a 2,600-clip dialect-balanced test set, with per-dialect tables) is documented on the source model card.

Limitations

• Offline mode only. Streaming (chunked encoder with cache tensors) is not exported.
• The language prompt is frozen to ar.
• 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

The base model nvidia/nemotron-3.5-asr-streaming-0.6b is governed by OpenMDW-1.1, which permits redistribution of derived materials when the license text and the notices of origin travel with them. The full text is in LICENSE.

The fine-tune's own model card states license: other without naming a licence, so the terms above are the ones that can be identified for these weights. Original fine-tune: Copyright oddadmix. Base model: Copyright NVIDIA Corporation.