Speech enhancement is a technical framework designed to improve degraded speech signals by suppressing noise, reverberation, and interference. It primarily addresses the problem of impaired speech intelligibility and naturalness caused by signal contamination in complex acoustic environments. This architecture broadly encompasses core tasks such as speech recovery, target speaker extraction, speech separation, and noise suppression, accurately preserving the acoustic features of the original speech while filtering out background noise. Research results show that modern speech enhancement effectively breaks through the performance bottlenecks of traditional statistical filters, significantly improving the human auditory experience and serving as a perfect complement to automatic speech recognition and audio understanding. In scenarios such as hearing aids, remote conferencing, and voice interaction in complex environments, it enables systems to achieve clear speech capture and parsing with extremely high robustness.

Speech enhancement technology originated from signal processing research in the late 1970s, with a key milestone being the 1979 paper by Steven F. Boll, a scholar at the University of Utah.Suppression of Acoustic Noise in Speech Using Spectral SubtractionIn his paper "(IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. 27, no. 2, pp. 113-120), he proposed to estimate the stationary background noise spectrum in real time using silent segments of speech and to directly subtract the noise component from the short-time Fourier amplitude spectrum of the noisy signal to achieve enhancement. This efficient spectral subtraction paradigm laid a practical foundation for subsequent research."

Building on this, in 1984 Y. Ephraim and D. Malah published another landmark paper, "..."Speech enhancement using a minimum-mean-square-error short-time spectral amplitude estimator(IEEE Trans. Acoust., Speech, Signal Process., vol. 32, no. 6, pp. 1109-1121). They assumed that the spectral components were Gaussian random variables and derived a short-time spectral amplitude estimator using the minimum mean square error criterion, effectively balancing noise suppression and speech distortion. In the mid-1990s, Ephraim and H.L. Van Trees further introduced the signal subspace method, decomposing the noisy signal into orthogonal signal and noise subspaces and performing projection filtering, improving performance in low signal-to-noise ratio environments.

Researchers from the University of Science and Technology of China (USTC) and Georgia Institute of Technology formally established a deep learning-based regression paradigm in September 2014. Representative research findings were published in a landmark paper in the field of audio and acoustic processing.A Regression Approach to Speech Enhancement Based on Deep Neural NetworksIn their work, they abstracted speech enhancement as a regression problem using deep neural networks, directly learning the implicit mapping from noisy spectra to clean spectra using DNNs. This broke through the limitations of traditional statistical assumptions and propelled the rapid development of speech enhancement in the AI era. These works collectively constitute a complete evolution from classical spectrum processing to data-driven intelligent systems, and have been widely applied in fields such as hearing aids, speech recognition, and real-time communication.