Main Result
This model is the accuracy tier. It beats Whisper-large-v3 on Czech formal/parliamentary speech and VoxPopuli in our evaluation matrix, while Whisper remains slightly ahead on FLEURS and Common Voice reference numbers.
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.
Table with columns: model, mode, ParCzech formal, FLEURS, VoxPopuli, Common Voice| model | mode | ParCzech formal | FLEURS | VoxPopuli | Common Voice |
|---|
openai/whisper-large-v3 | offline seq2seq reference | 5.76 | 12.10 | 13.70 | 11.70 |
inferRouter/qwen3-asr-cs-1.7b | offline AED | 3.43 | 12.37 | 9.79 | 11.76 |
inferRouter/nemotron-cs-asr-0.6b | streaming RNNT | 5.87 | 17.18 | 11.28 | 14.28 |
| internal Qwen3-ASR 0.6B full FT | offline AED | 11.84 | 24.37 | 16.34 | 21.32 |
Interpretation
- Qwen 1.7B is the best model in this portfolio for offline / batch Czech ASR.
- It clearly beats Whisper-large-v3 on ParCzech formal and VoxPopuli.
- It is near Whisper-large-v3 on FLEURS/Common Voice but does not robustly beat Whisper there.
- Nemotron v4 is less accurate but is the true streaming / low-latency tier.
- The internal Qwen3-ASR 0.6B full fine-tune failed as a challenger and is not published.
Portfolio Position
Table with columns: model, best use, strength, tradeoff| model | best use | strength | tradeoff |
|---|
inferRouter/qwen3-asr-cs-1.7b | offline / batch Czech ASR | best Czech formal/legal-adjacent accuracy, strong formatting | autoregressive AED; not true streaming |
inferRouter/nemotron-cs-asr-0.6b | live / streaming Czech ASR | RNNT streaming, RTFx ~164, lower latency | lower WER ceiling |
openai/whisper-large-v3 | external offline baseline / teacher | strong multilingual general ASR |
Speed / Deployment Notes
- Batch-1 decode on RTX PRO 6000 Blackwell: RTFx ~3.1, mean latency around 2-3.5 s on measured short/medium clips.
- Batched offline throughput is much higher: dynamic batched eval reached roughly RTFx ~130 on the formal board.
- This is not a true streaming model. It is intended for record-then-transcribe or offline/batch transcription.
Training
- Base:
Qwen/Qwen3-ASR-1.7B-hf.
- Q1b warm-start: intermediate Czech LoRA adaptation.
- Q1c: Czech full-anchor adaptation.
- Selected checkpoint: Q1c
ckpt-85000 (best formal + low slot error among floor checkpoints), merged into the base model.
Detailed training-corpus composition is intentionally not listed in this public model card. Public benchmark test splits were held out from training.
Usage
import torch
from transformers import AutoProcessor, Qwen3ASRForConditionalGeneration
model_id = "inferRouter/qwen3-asr-cs-1.7b"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen3ASRForConditionalGeneration.from_pretrained(
model_id,
torch_dtype=torch.bfloat16,
device_map="cuda",
)
Intended Use
Offline Czech ASR for high-accuracy transcription, especially formal/parliamentary/legal-adjacent speech. For live captioning or low-latency streaming, use the Nemotron RNNT model instead.
Licence Notes
Base model is Apache-2.0. The merged model is released under Apache-2.0. Review the model licence and your downstream use case before redistribution or commercial deployment.