I am a Ph.D. student in the Graduate Institute of Communication Engineering at National Taiwan University, advised by Prof. Hung-yi Lee.
My research focuses on speech and language processing, especially instruction-aware speech retrieval, large audio-language models, and self-supervised speech representations.
I received my B.S. in Computer Science and Information Engineering from NTU in 2023. During my undergraduate studies, I worked with Prof. Hung-yi Lee and Prof. Yun-Nung (Vivian) Chen. I also completed research internships at ASUS and the Institute of Information Science, Academia Sinica.
Instruction-aware retrieval of speech content and attributes.
Audio-text interaction, evaluation, and robustness.
Self-supervised learning and prosody-related tasks.
New paper alert: IRIS-Embed is now on arXiv.
Our Long Paper INSPIRE is accepted to Interspeech 2026 in Sydney, Australia .
I entered the Ph.D. program at NTU through direct admission, advised by Prof. Hung-yi Lee.
Our paper When Silence Matters is accepted to ICASSP 2026 in Barcelona, Spain .
Our paper Transferring Textual Preferences to Vision-Language Understanding through Model Merging is accepted to ACL 2025 in Vienna, Austria .
arXiv:2611.09122
IRIS-Embed combines contrastive training and instruction-aware audio layer weighting to improve speech retrieval and performance on general audio tasks.
Interspeech 2026 · Long Paper
INSPIRE evaluates speech retrieval guided by instructions about content, speaker identity, speaking style, background sounds, or combinations of these attributes.
Interspeech 2026
This study examines how answer order affects large audio-language models and explores permutation-based strategies for more reliable evaluation.
ICASSP 2026
This study shows that irrelevant audio, including silence and noise, can disrupt text reasoning in large audio-language models.
ACL 2025
This method merges text reward models with vision-language models to transfer textual preferences and improve multimodal evaluation without additional training.
EMNLP 2024
DogeRM merges domain-specific language models with general reward models to improve reward modeling without additional domain-specific preference data.
Findings of ACL 2023
This method combines generated visual cues with text and image encoders to improve phrase understanding and entity clustering.