So Yeong Park

So Yeong Park

Undergraduate Researcher
AINC Lab
Hanshin University

Evaluation and inference parameter analysis of Qwen3-ASR on multiple Korean speech corpora

다중 한국어 음성 코퍼스에서의 Qwen3-ASR 평가 및 추론 파라미터 분석

So Yeong Park, Seon Man Kim

The Journal of the Acoustical Society of Korea, 45(5), 582–590

Journal article · 2026

Paper (PDF)Publication record

Figure 3: Qwen3-ASR 0.6B and 1.7B have lower character error rates than Whisper medium and large-v3, respectively, on KsponSpeech eval_clean and eval_other.
Fig. 3 from the paper. Character error rate with spaces removed (CERNS) on KsponSpeech eval_clean and eval_other; each model uses its best tested decoding setting on these two sets. Lower is better.

We evaluate Qwen3-ASR 0.6B and 1.7B without fine-tuning on ten evaluation sets from nine Korean speech corpora. The study examines the effects of inference settings and identifies a practical default for the tested Korean speech conditions.

All publications