SE-AGCNet unifies speech enhancement and loudness control for meetings

Conventional audio pipelines treat speech enhancement (SE) and automatic gain control (AGC) as separate modules, leading to a performance trade-off: applying AGC before SE amplifies background noise, while prioritizing SE over-suppresses low-volume speech. To overcome this, researchers Jinming Zhang, Wei Rao, Xionghu Zhong, and Eng Siong Chng propose SE-AGCNet, an end-to-end framework that jointly optimizes both tasks. Tailored for meeting scenarios with significant volume variations, SE-AGCNet leverages the synergy between SE and AGC—SE preserves quiet speech, enabling the AGC component to adjust volume effectively without distortion.

In addition to the core architecture, the team introduces a specialized data simulation pipeline called SE-AGC-DataGen to generate realistic training data. They also incorporate standardized loudness evaluation metrics: integrated loudness (LUFS), short-term loudness (St LUFS), and loudness range (LRA). Experiments demonstrate that SE-AGCNet consistently achieves target loudness while improving both speech quality and automatic speech recognition (ASR) accuracy over competitive baselines. The work has been accepted at Interspeech 2026, signaling strong peer validation. This unified approach is particularly valuable for remote meetings, voice assistants, and teleconferencing systems where consistent volume and clarity are critical.