I am a second-year Ph.D. candidate at KAIST advised by Professor Joon Son Chung, and I earned my M.S. from KAIST. My research focuses on audio representation learning and its connection to Audio-LLM systems. I am interested in how audio representations can be made robust under real-world variability, transformed across representation spaces, and exposed to language-model interfaces.
My work started from robust speech and speaker representation learning, where audio models must preserve task-relevant information while handling language, environment, session, channel, and recording mismatch. Building on this foundation, I have recently explored generative representation alignment: using diffusion-based methods to enhancement or transform audio representations at the latent level. In this direction, SEED studies speaker embedding enhancement, while Diffusion-Link bridges the audio-text modality gap by transporting audio representations toward text-aligned spaces for Audio-LLMs and multimodal LLMs.
More recently, I have been interested in specialized Audio-LLM systems. As voice becomes a primary interface for robots, smart glasses, wearables, vehicles, and other physical AI systems, Audio-LLMs will need capabilities beyond ASR. My recent work SpeakerLLM explores this direction by enabling an Audio-LLM to understand speaker identity, recording conditions, utterance-pair relations, and verification reasoning through a natural-language interface.
Going forward, I am interested in building audio-native intelligence for voice-first systems, including speaker-aware Audio-LLMs, audio-text representation alignment, and full-duplex speech-to-speech agents. My long-term goal is to build Audio-LLM systems that use audio not only as transcribed text, but as rich evidence for reasoning, personalization, and real-time interaction.
Advisor: Joon Son Chung (Multimodal AI Lab)
Advisor: Joon Son Chung (Multimodal AI Lab)
