inferRouter/nemotron-cs-asr-0.6b

Nemotron 3.5 ASR Streaming 0.6B - Czech (v4)

Streaming Czech ASR fine-tuned from nvidia/nemotron-3.5-asr-streaming-0.6b (EncDecRNNTBPEModelWithPrompt, RNNT, att_context_size=[56,0] ~80 ms lookahead).

This is the streaming / production-latency model in the InferRouter Czech ASR portfolio. For maximum offline accuracy, use inferRouter/qwen3-asr-cs-1.7b.

Main Result

v4 improves over both the public Nemotron base and our previous Czech v3 on every Czech eval split while staying fully streaming.

WER is word-level, lowercased, punctuation-insensitive, and computed with the same frozen Czech evaluation board where locally measured. Whisper-large-v3 formal was measured locally on the same ParCzech formal split; the Whisper FLEURS/VoxPopuli/Common Voice values are included as external reference values.

model	mode	ParCzech formal	FLEURS	VoxPopuli	Common Voice
nvidia/nemotron-3.5-asr-streaming-0.6b base	streaming RNNT	23.8	25.0	26.5	24.6
InferRouter Nemotron v3	streaming RNNT	9.04	18.76	14.82	16.55
openai/whisper-large-v3	offline seq2seq reference	5.76	12.10	13.70	11.70
this model: InferRouter Nemotron v4	streaming RNNT	5.87	17.18	11.28	14.28
inferRouter/qwen3-asr-cs-1.7b	offline AED	3.43	12.37	9.79	11.76
internal Qwen3-ASR 0.6B full FT	offline AED	11.84	24.37	16.34	21.32

Interpretation

• Nemotron v4 is the best streaming Czech ASR model in this portfolio.
• It is approximately at Whisper-large-v3 formal accuracy while remaining RNNT streaming.
• Qwen 1.7B is more accurate, but it is offline / autoregressive AED and not the low-latency streaming tier.
• The internal Qwen3-ASR 0.6B full fine-tune is dominated by Nemotron v4 and is not published.

Portfolio Position

model	best use	strength	tradeoff
inferRouter/nemotron-cs-asr-0.6b	live / streaming Czech ASR	RNNT streaming, RTFx ~164, low latency	lower WER ceiling than Qwen 1.7B
inferRouter/qwen3-asr-cs-1.7b	offline / batch Czech ASR	best accuracy, formal WER 3.43	autoregressive AED, not true streaming
openai/whisper-large-v3	external offline baseline / teacher	strong multilingual general ASR	not the InferRouter streaming tier

Training

• Warm-start: InferRouter Nemotron v3.
• Recipe: Czech RNNT SFT, bf16, batch 24 on RTX PRO 6000 Blackwell, 90k steps, best checkpoint by dev WER.
• Best dev formal val_wer: 0.0719.

Detailed training-corpus composition is intentionally not listed in this public model card. Public benchmark test splits were held out from training.

Usage (NeMo)

from nemo.collections.asr.models import ASRModel

model = ASRModel.restore_from("nemotron-cs-asr-0.6b.nemo", map_location="cuda")
model.change_attention_model(self_attention_model="rel_pos_local_attn", att_context_size=[56, 0])
print(model.transcribe(["utterance.wav"], language="cs")[0])

Intended Use

Czech streaming ASR: dictation, live captioning, low-latency transcription and server-side streaming. Formal/parliamentary/legal-adjacent Czech is the strongest domain; general/read speech works with higher WER.

Licence Notes

Weights are released under CC-BY-4.0. Review the model licence and your downstream use case before redistribution or commercial deployment.