Electronic Engineering Research Team Led by Professor Park Hyung-min Has a Paper Accepted at “ICML 2026,” the World’s Premier AI Conference
A research team from the Intelligent Information Processing Laboratory, led by Professor Park Hyung-min of the Department of Electronic Engineering at Sogang University, will present a paper at the International Conference on Machine Learning (ICML 2026), the world’s most prestigious conference on artificial intelligence and machine learning.
This research was led by Dr. Shin Eui-hyup with the participation of Ph.D. student Ko Jae-hyun and Master’s student Jeong Woo-cheol from the Department of Electronic Engineering. The title of the presented paper is “Query-Based Asymmetric Modeling with Decoupled Input-Output Rates for Speech Restoration.” The research team proposed a new deep learning network architecture capable of effectively restoring input speech recorded in poor environments at various sampling frequencies.
Speech restoration is a technology that restores speech whose quality has been degraded due to noise, reverberation, bandwidth limitations, or digital distortion to a state that sounds clear and natural to the human ear. Previous studies had the limitation of designing models based on the assumption of fixed input and output sampling frequencies.
To address this issue, the research team designed a more flexible speech restoration model capable of processing speech with arbitrary input sampling frequencies while reconstructing it at any desired output sampling frequency specified by the user. To achieve this, they proposed an asymmetric network architecture based on an input-band analysis encoder and an output-band restoration decoder, which separates the roles of the encoder (analyzing the input speech’s frequency band information) and the decoder (restoring the output speech).
Meanwhile, ICML is one of the world’s leading international conferences covering advancements in artificial intelligence (AI) and machine learning. Every year, researchers and industry professionals from academia and industry around the world gather to present and discuss the latest research findings and innovative technologies. It is particularly regarded as a premier conference where major research in core AI fields—such as machine learning theory, algorithms, deep learning, and reinforcement learning—is shared. ICML 2026 will be held at COEX in Seoul, South Korea, from July 6 to 11, 2026.
▶논문제목: Query-Based Asymmetric Modeling with Decoupled Input-Output Rates for Speech RestoratiPaper Title: Query-Based Asymmetric Modeling with Decoupled Input-Output Rates for Speech Restoration
▶저자 정보: 신의협(제1저자), 고재현(제2저자), 정우철(제3저자), 박형민 교수(교신저자, 서강대)Authors: Shin Eui-hyup (First Author), Ko Jae-hyun (Second Author), Jeong Woo-cheol (Third Author), and Professor Park Hyung-min (Corresponding Author, Sogang University)
▶Demo Page: https://tf-restormer.github.io/demo/
[SEO Keyword]
ICML 2026, Speech Restoration Technology, Asymmetric Network Architecture
This research was led by Dr. Shin Eui-hyup with the participation of Ph.D. student Ko Jae-hyun and Master’s student Jeong Woo-cheol from the Department of Electronic Engineering. The title of the presented paper is “Query-Based Asymmetric Modeling with Decoupled Input-Output Rates for Speech Restoration.” The research team proposed a new deep learning network architecture capable of effectively restoring input speech recorded in poor environments at various sampling frequencies.
Speech restoration is a technology that restores speech whose quality has been degraded due to noise, reverberation, bandwidth limitations, or digital distortion to a state that sounds clear and natural to the human ear. Previous studies had the limitation of designing models based on the assumption of fixed input and output sampling frequencies.
To address this issue, the research team designed a more flexible speech restoration model capable of processing speech with arbitrary input sampling frequencies while reconstructing it at any desired output sampling frequency specified by the user. To achieve this, they proposed an asymmetric network architecture based on an input-band analysis encoder and an output-band restoration decoder, which separates the roles of the encoder (analyzing the input speech’s frequency band information) and the decoder (restoring the output speech).
Meanwhile, ICML is one of the world’s leading international conferences covering advancements in artificial intelligence (AI) and machine learning. Every year, researchers and industry professionals from academia and industry around the world gather to present and discuss the latest research findings and innovative technologies. It is particularly regarded as a premier conference where major research in core AI fields—such as machine learning theory, algorithms, deep learning, and reinforcement learning—is shared. ICML 2026 will be held at COEX in Seoul, South Korea, from July 6 to 11, 2026.
▶논문제목: Query-Based Asymmetric Modeling with Decoupled Input-Output Rates for Speech RestoratiPaper Title: Query-Based Asymmetric Modeling with Decoupled Input-Output Rates for Speech Restoration
▶저자 정보: 신의협(제1저자), 고재현(제2저자), 정우철(제3저자), 박형민 교수(교신저자, 서강대)Authors: Shin Eui-hyup (First Author), Ko Jae-hyun (Second Author), Jeong Woo-cheol (Third Author), and Professor Park Hyung-min (Corresponding Author, Sogang University)
▶Demo Page: https://tf-restormer.github.io/demo/
[SEO Keyword]
ICML 2026, Speech Restoration Technology, Asymmetric Network Architecture