Modelpublic

notmax123/nemotron-3.5-asr-hebrew-streaming-0.6b

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

1likes23downloads

Model Card

Nemotron 3.5 ASR Hebrew Streaming 0.6B

Fine-tuned nvidia/nemotron-3.5-asr-streaming-0.6b for Hebrew (he-IL) cache-aware streaming ASR.

Model details


Base model	`nvidia/nemotron-3.5-asr-streaming-0.6b`
Architecture	FastConformer-Transducer (RNNT) with langID prompt conditioning
Language	Hebrew (`he-IL`, `prompt_mode=langID`)
Sample rate	16 kHz
Streaming	Cache-aware; deployment latency ~80 ms (`att_context_size=[56,0]`)
Dev val WER	17.41% (VoxKnesset hold-out dev)
Format	NeMo `.nemo` checkpoint (~2.4 GB)

Training

Fine-tuned following the NVIDIA Nemotron 3.5 ASR fine-tuning guide on Hebrew speech data (VoxKnesset, synthetic TTS, podcasts, ivrit.ai sources). Training uses Lhotse bucketing, bf16, NoamAnnealing, and telephony-style augmentations (noise, gain, SpecAugment).

Usage

NeMo (recommended)

python

import nemo.collections.asr as nemo_asr

model = nemo_asr.models.ASRModel.restore_from(
    "nemotron-3.5-asr-hebrew-streaming-0.6b.nemo"
)

# Offline transcription with Hebrew langID prompt
hypotheses = model.transcribe(
    ["audio.wav"],
    batch_size=1,
    verbose=False,
)
print(hypotheses[0].text)

For streaming inference, use cache-aware streaming with att_context_size=[56, 0] as in the base model card.

Download

bash

hf download notmax123/nemotron-3.5-asr-hebrew-streaming-0.6b

Limitations

—Optimized for Hebrew; other languages are not supported by this checkpoint.
—Long audio should be chunked for streaming evaluation (see project eval scripts).
—Intended for research and production ASR; verify on your domain before deployment.

License

Apache 2.0 (same as the base Nemotron 3.5 ASR model).