Research.

NAACL

Automatic, sentiment-guided framework that produces large-scale preference datasets without human annotation — improves emotional appropriateness in customer-support LLMs.

ACL-W

Disentangles decoder representations and dynamically fuses task-specific information — state of the art on the ESConv benchmark for both strategy prediction and response generation.

ICASSPW

Fine-tuning strategy for hearing-aid–oriented source separation. Softening targets with previous-epoch predictions gives +1.2 dB SDR over the baseline.

KCC

LLM-based prompt augmentation that explicitly models ASMR triggers — produces more detailed sound descriptions and improves T2A quality.

ISMIR

Predicts the track role of single-instrument sequences automatically. 87% symbolic / 84% audio accuracy — reduces manual annotation in MIR pipelines.

TOOL

// PROJECT

Renders MIDI into audio using NSynth instrument samples, holding one instrument preset consistent across an entire sequence — used to build large symbolic-to-audio datasets.

TOOL

// PROJECT

Splits a multi-track MIDI file into individual instrument tracks, extracts per-track metadata, and supports measure-based truncation and saving.

TOOL

// PROJECT

Soothing-sound audio generator with five distinct noise colors and three weighting filters. Copyrighted with the Korea Copyright Commission.