Modelpublic

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

sourceHugging Faceotherupdated 2mo agoView on Hugging Face

0likes947downloads

Model Card

Nemotron 3.5 ASR Streaming — Hebrew Fine-tune

Fine-tuned from `nvidia/nemotron-3.5-asr-streaming-0.6b` on ivrit.ai Hebrew datasets.

Evaluation

Metric	Base Model	Fine-tuned
Raw WER	0.4857	0.4260
Normalized WER	0.3943	0.3324
CER	—	0.1807

Relative WER improvement: 12.3%

Training details

—Freeze mode: encoder
—Schedule: CosineAnnealing (peak LR=3e-5, min_lr=1e-6, warmup 2%)
—Epochs: 3
—Batch duration: 15s
—Grad accumulation: 8
—Precision: bf16 mixed
—GPU: NVIDIA L4

Model Hash

SHA-256: 8d588c7220818729f23d7faf3c2d87e91195c83dc2d0dbcf0b9e025c445b5d55

Usage

python

import nemo.collections.asr as nemo_asr
model = nemo_asr.models.ASRModel.restore_from(nemotron-hebrew-best.nemo)
transcriptions = model.transcribe([hebrew_audio.wav])

Set target_lang=he-IL for streaming.