Interspeech 2026

SE-AGCNet

Jinming Zhang1 Wei Rao2 Xionghu Zhong3 Eng Siong Chng2

1Zhejiang University, China 2Nanyang Technological University, Singapore 3Hunan University, China

An end-to-end framework for joint speech enhancement and loudness control in meeting scenarios with large volume variations.

Code Audio Demos Method

Abstract

Conventional audio pipelines usually treat speech enhancement and automatic gain control as separate modules. This can amplify background noise when AGC is applied before enhancement, or suppress quiet speech when enhancement is applied before AGC. SE-AGCNet jointly optimizes both tasks end-to-end for meeting scenarios, preserving low-volume speech while controlling loudness. The project also includes SE-AGC-DataGen and loudness-aware evaluation with integrated LUFS, short-term LUFS, and LRA.

Joint SE + AGC

End-to-end One framework jointly optimizes speech enhancement and automatic gain control.
Meeting focused Designed for far-field speech, noise, reverberation, and strong speaker loudness variation.
Loudness aware Evaluated with LUFS, short-term LUFS, LRA, speech quality, and ASR metrics.
SE-AGCNet pipeline diagram

Audio Demos

Each row compares the noisy input, MP-SENet baselines, their pyagc post-processing variants, and SE-AGCNet.

LibriAGC

Input
MP-SENet (Orig)
MP-SENet (Orig) + pyagc
MP-SENet (SE)
MP-SENet (SE) + pyagc
MP-SENet (AGC)
SE-AGCNet

MMCSG

Input
MP-SENet (Orig)
MP-SENet (Orig) + pyagc
MP-SENet (SE)
MP-SENet (SE) + pyagc
MP-SENet (AGC)
SE-AGCNet

AliMeeting Far-field

Input
MP-SENet (Orig)
MP-SENet (Orig) + pyagc
MP-SENet (SE)
MP-SENet (SE) + pyagc
MP-SENet (AGC)
SE-AGCNet

Resources

Code, training scripts, inference scripts, pyagc, and data generation utilities are available in the GitHub repository. Pre-trained model and dataset links are documented in the repository README.