Speech Enhancement Overview

Principles

Definition

Speech enhancement is a research field and set of signal processing techniques focused on improving the perceptual quality and intelligibility of speech signals degraded by acoustic noise, reverberation, or other distortions. It investigates algorithmic methods to suppress unwanted components and restore desired speech characteristics, critical for human-machine interaction and effective communication in adverse acoustic environments.

Ontological type

Core Algorithms

Objective Evaluation

Application Domains

Key developments

Key Figures

Probabilistic Model-Based Enhancement era

N. Morgan [1] was active at University of California, Berkeley [3] during the 1991-1997 era. His notable contribution was the paper 'RASTA processing of speech' [6], which introduced robust spectral-temporal processing that improved intelligibility under adverse conditions. Y. Ephraim [2] held positions at MIT [4] and University of Patras [5] during the era. His contributions include 'Statistical-model-based speech enhancement systems' [7], 'A Bayesian estimation approach for speech enhancement using hidden Markov models' [8], and 'Speech enhancement using state dependent dynamical system model' [9], collectively establishing probabilistic estimators and state-space models as core tools for the era.

Data-Driven Spectral Enhancement era

Richard C. Hendriks [1] was associated with Delft University of Technology [3] during the Data-Driven Spectral Enhancement era. His key contributions include the paper An Algorithm for Intelligibility Prediction of Time–Frequency Weighted Noisy Speech [6] and the paper Unbiased MMSE-Based Noise Power Estimation With Low Complexity and Low Tracking Delay [7], which provided practical objective-driven tools to assess and improve spectral enhancement in nonstationary noise. Yi Hu [2] conducted work across Shandong University [4] and the University of Calgary [5] in this era. Her contributions include evaluating objective quality measures for speech enhancement [8], proposing a generalized subspace approach for enhancing speech corrupted by colored noise [9], and the algorithm that improves speech intelligibility in noise for normal-hearing listeners [10], collectively advancing objective assessment, multichannel/subspace processing, and intelligibility-focused enhancement that defined data-driven spectral methods.

End-to-End Time-Phase Enhancement era

DeLiang Wang[1] stands as a leading figure in speech enhancement, with affiliations including The Ohio State University[3] and Northwestern Polytechnical University[4] during this era. Wang[1] contributed to end-to-end time-domain speech enhancement with TCNN[7], Learning Complex Spectral Mapping With Gated Convolutional Recurrent Networks for Monaural Speech Enhancement[8], and A New Framework for CNN-Based Speech Enhancement in the Time Domain[9], advancing real-time processing and direct modeling of phase information in this era. Dong Yu[2] contributed to multitalker speech separation with utterance-level permutation invariant training from Johns Hopkins University[5] and Nanyang Technological University[6], notably through Multitalker Speech Separation With Utterance-Level Permutation Invariant Training of Deep Recurrent Neural Networks[10]. This work addressed permutation ambiguity in multi-speaker scenarios, enabling stable end-to-end training and informing the development of end-to-end time-phase enhancement models that integrate multi-talker capabilities.