Nemotron 3.5 ASR – Fine-tuned on CV24

Fine-tuned from nvidia/nemotron-3.5-asr-streaming-0.6b on the Mozilla common-voice-scripted-speech-24-0-hungarian dataset using NVIDIA NeMo.

Usage

import nemo.collections.asr as nemo_asr
from huggingface_hub import hf_hub_download

nemo_path = hf_hub_download(
    repo_id="GaborMadarasz/nemotron-3.5-asr-finetuned-CV24",
    filename="FastConformer-Transducer-BPE-Prompt-Streaming.nemo",
    local_dir=".",
    local_dir_use_symlinks=False,
)
model = nemo_asr.models.EncDecRNNTBPEModel.restore_from(nemo_path)
transcriptions = model.transcribe(["audio.wav"])
print(transcriptions)

Training details

• Base model: nvidia/nemotron-3.5-asr-streaming-0.6b
• Framework: NVIDIA NeMo 26.06+
• Precision: fp32
• Optimizer: AdamW lr=1e-4
• Dataset: Mozilla Common Voice 24 (Hungarian)