Research
Research at AINC Lab, Hanshin University, advised by Prof. Seon Man Kim.
Evaluating Qwen3-ASR for Korean Speech Recognition
2026 · Journal accepted
Evaluated how model size and inference settings affect recognition accuracy across Korean speech datasets.
- Compared Qwen3-ASR 0.6B and 1.7B on 10 Korean evaluation sets.
- Analyzed language settings, maximum generation length, temperature, and repetition penalty using character error rate.
- Designed the study, implemented experiments, analyzed results, and wrote the paper.
DCASE 2026 Task 6 · Audio Moment Retrieval
2026 — Present · Ongoing
Studying how to locate audio segments that match natural-language queries.
- Reproduced the official baseline and its full evaluation pipeline.
- Experimenting with MS-CLAP audio-language representations and DETR-based temporal localization models.
Academic-Regulation QA Data & Error-Correction Training
2025 — 2026 · Conference awards
Built and evaluated training data for question answering grounded in university academic regulations.
- Designed a rubric-based dataset for academic-regulation questions and answers in 2025.
- Designed and reviewed 9,958 synthetic examples labeled PASS, FAIL, and OOS; fine-tuned Qwen2.5-7B-Instruct and Llama-3.1-8B with 4-bit QLoRA.
- Measured a FAIL correction rate of 0.9897 when ground-truth review signals were supplied; identified error review as a bottleneck in end-to-end inference.