Downloads · 30 days
19
42% of all-time downloads
jinhwan000/qwen3-asr-1.7b-ko-beta-ov-int8
qwen3-asr-1.7b-ko-beta-ov-int8 is a automatic speech recognition model from jinhwan000. Use it when you need speech turned into text. The card lists the license as apache-2.0.
OpenVINO IR export of TeamUNIVA/qwen3asr1.7bkobeta (a Korean full fine-tune of Qwen/Qwen3-ASR-1.7B) with INT8 weight compression (NNCF INT8SYM), for Intel CPU/iGPU/dGPU inference. Same layout and conversion recipe as…
Downloads · 30 days
19
42% of all-time downloads
All-time downloads
45
Public
Repo size
3 GB
Likes
0
Public
Click a slice to open those files.
.bin3 GB · 99%
From the Hugging Face model README
OpenVINO IR export of TeamUNIVA/qwen3_asr_1.7b_ko_beta (a Korean full fine-tune of Qwen/Qwen3-ASR-1.7B) with INT8 weight compression (NNCF INT8_SYM), for Intel CPU/iGPU/dGPU inference. Same layout and conversion recipe as jinhwan000/qwen3-asr-1.7b-ov-int8 (the base-model export), so the two are drop-in interchangeable.
chat_template.jinja is copied from the base model. The TeamUNIVA repository ships no chat template, which makes processor.apply_chat_template fail; the tokenizer (vocab/merges) is identical to the base model, so the base template is used verbatim.Measured against the base-model export on the same audio, with the same serving code and knobs (RSttStreamer streaming pipeline: 2.0 s chunks, 2-tier unfixed tokens, hallucination fallback). Test set: 39 VAD-segmented utterances from one real Korean call-center call (78 s, 16 kHz), scored on the 37 utterances with a non-empty reference (410 characters). "numag" = character error rate that ignores Arabic-vs-Korean numeral notation differences (both readings are tried and the better one is taken).
Intel Arc B580, OpenVINO 2026.2, INT8, paged decoder:
| model | system-prompt context | batch CER | batch numag | streaming CER | streaming numag | batch RTF |
|---|---|---|---|---|---|---|
| base (Qwen3-ASR-1.7B INT8) | yes | 4.15% | 4.15% | 5.85% | 4.15% | 0.112 |
| this model | yes | 10.00% | 5.61% | 9.76% | 5.37% | 0.113 |
| base (Qwen3-ASR-1.7B INT8) | no | 7.80% | 7.32% | 7.32% | 6.83% | 0.109 |
| this model | no | 11.22% | 6.10% | 11.71% | 6.34% | 0.110 |
The same pattern was reproduced with vLLM bf16 on an RTX 5080 (base 3.90% vs this model 6.10% numag with context; 7.07% vs 5.37% without).
What this means in practice:
781104, 26년 5월, 2천) far more often than with the base model. If your transcription convention uses Korean number words, expect a large raw-CER penalty.Sample size is small (one call, 410 characters), so treat the absolute numbers as indicative, not definitive.
config.json preprocessor_config.json tokenizer.json tokenizer_config.json vocab.json merges.txt
added_tokens.json special_tokens_map.json chat_template.jinja hann_400.f32 mel_filters.f32
thinker/openvino_thinker_embedding_model.{xml,bin}
thinker/openvino_thinker_audio_model.{xml,bin}
thinker/openvino_thinker_audio_encoder_model.{xml,bin}
thinker/openvino_thinker_language_model.{xml,bin}
Apache-2.0, following the upstream model and the base model.