Slovenian streaming ASR — Nemotron-3.5 fine-tune (v3)
Trained by MediaAtlas — LLM fine-tuning on your own data, trained in the EU, weights delivered. Pricing · All our models
nvidia/nemotron-3.5-asr-streaming-0.6b (FastConformer-CacheAware RNNT, 638M) fine-tuned for
Slovenian on ARTUR (858 h) + Common Voice — a low-latency streaming model that works
across read and spontaneous speech. Runs on GPU via NeMo or on CPU via parakeet.cpp (GGUF).
Results (WER, normalized, held-out)
| Test set | WER |
|---|---|
| Common Voice sl | 25.07% |
| ARTUR multi-domain | 23.42% |
| ARTUR parliamentary (spontaneous) | 21.66% |
| FLEURS sl | 33.07% |
Comparison — vs yuriyvnv/parakeet-tdt-0.6b-slovenian
Different design goals; honest numbers (normalized WER):
| Test set | This model (streaming) | parakeet-tdt-0.6b-slovenian (non-streaming) |
|---|---|---|
| FLEURS sl | 33.07% | 17.74% |
| Common Voice sl | 25.07% | 8.81% |
| ARTUR parliamentary (spontaneous) | 21.66% | not reported |
parakeet-tdtis better on clean read speech — stronger non-streaming base (parakeet-tdt-0.6b-v3), trained on read data (Common Voice + synthetic TTS).- This model is a streaming (real-time, low-latency, CPU-deployable) model trained on 858 h ARTUR + Common Voice, so it stays robust on spontaneous/conversational speech (parliamentary ~22%) — a domain read-only models don't target. (CV test differs slightly: CV26 here vs CV17 there.)
Pick by use case: read-speech accuracy → parakeet-tdt; real-time streaming + spontaneous speech → this model.
Files
nemotron_sl.nemo— NeMo checkpoint (GPU inference / further fine-tuning)asr_sl_v3.gguf(q8_0) — parakeet.cpp CPU inference
How to run
# CPU, streaming, via parakeet.cpp — use lang tag sl-SI, 16 kHz mono
parakeet-cli --model asr_sl_v3.gguf --lang sl-SI audio.wav
# GPU via NeMo
from nemo.collections.asr.models import ASRModel
m = ASRModel.restore_from("nemotron_sl.nemo")
m.transcribe(["audio.wav"]) # 16 kHz mono; manifest lang/target_lang = "sl-SI"
License & data provenance
Released CC BY-SA 4.0 (to honor the ARTUR ShareAlike terms).
| Input | License |
|---|---|
Base nvidia/nemotron-3.5-asr-streaming-0.6b | OpenMDW-1.1 (retain NOTICE) |
| ARTUR 1.0 (CLARIN.SI, 858 h) | CC BY-SA 4.0 (attribution + ShareAlike) |
| Common Voice sl | CC0 |
Credit CLARIN.SI (ARTUR: hdl.handle.net/11356/1776 audio, /11356/1772 transcriptions), NVIDIA, and Mozilla Common Voice.
Citation
Cite this work (Tadej Fius, MediaAtlas Ltd):
@misc{fius2026asr,
title = {Nemotron-3.5 Slovenian Streaming ASR},
author = {Fius, Tadej},
year = {2026},
publisher = {MediaAtlas Ltd},
howpublished = {Hugging Face},
url = {https://huggingface.co/texdata/nemotron-3.5-asr-streaming-slovenian}
}
Upstream / source citations:
@misc{nemotron_asr, title={Nemotron-3.5 ASR Streaming 0.6B (FastConformer-CacheAware RNNT)}, author={{NVIDIA}}, url={https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b}}
@article{kuchaiev2019nemo, title={NeMo: a toolkit for building AI applications using Neural Modules}, author={Kuchaiev, O. and others}, journal={arXiv:1909.09577}, year={2019}}
@misc{artur, title={ARTUR: Speech Database of Slovene}, author={{CLARIN.SI}}, note={audio hdl:11356/1776, transcriptions hdl:11356/1772}, url={http://hdl.handle.net/11356/1776}}
Base under OpenMDW-1.1 (retain NOTICE). Data: ARTUR (CLARIN.SI); credit NVIDIA + CLARIN.SI.