Speech Recognition Overview

Principles

Definition

Speech recognition is an interdisciplinary research field and computational methodology concerned with automatically transcribing spoken language into textual or other structured data. This process involves analyzing acoustic speech signals to identify phonetic and linguistic content, enabling machines to understand and process human voice input.

Ontological type

Core Algorithms

Acoustic Features

Language Modeling

Key developments

Key Figures

Probabilistic Modeling and Search era

Fan-Gang Zeng [1], affiliated with Syracuse University [3], exemplified key developments in the Probabilistic Modeling and Search era. His 1995 paper Speech Recognition with Primarily Temporal Cues [5] introduced a temporal-cue emphasis that enhanced the robustness of likelihood-based decoders, reflecting the era’s emphasis on probabilistic sequence modeling and search. N. Morgan [2], associated with University of California, Berkeley [4], advanced feature processing and neural approaches through works such as RASTA processing of speech [6] and RASTA-PLP speech analysis technique [7]. The CDNN: a context dependent neural network for continuous speech recognition [8] further illustrated how context-aware neural models could be integrated with probabilistic decoding to manage variability in speech.

Feature-space Adaptation era

Douglas A. Reynolds [1] is a prominent researcher associated with the Massachusetts Institute of Technology [3]. His key contributions include Speaker Verification Using Adapted Gaussian Mixture Models [4] and The NIST speaker recognition evaluation – Overview, methodology, systems, results, perspective [5], which collectively advanced cross-speaker robustness and the standardization of large-vocabulary decoding under nonideal conditions. Thomas F. Quatieri [2] is a researcher associated with the Massachusetts Institute of Technology [3]. Quatieri's key contribution in this area is Speaker Verification Using Adapted Gaussian Mixture Models [4], illustrating how adaptation-based feature-space techniques supported robust verification under nonideal conditions.

End-to-End Self-Supervised era

Dong Yu [1] is a leading figure in end-to-end speech modeling during this era, with affiliations at Johns Hopkins University [3] and the University of Washington [4]. His contributions, exemplified by Deep Neural Networks for Acoustic Modeling in Speech Recognition: The Shared Views of Four Research Groups [6], Convolutional Neural Networks for Speech Recognition [7], and Deep Neural Networks for Acoustic Modeling in Speech Recognition [8], helped establish deep networks as the core of acoustic modeling and demonstrated data-efficient architectures crucial for scaling ASR in this era. Li Deng [2] played a pivotal role at Stanford University [5] and the University of Washington [4] in shaping end-to-end learning for speech. His work, reflected in the same foundational papers as above—Deep Neural Networks for Acoustic Modeling in Speech Recognition: The Shared Views of Four Research Groups [6], Convolutional Neural Networks for Speech Recognition [7], and Deep Neural Networks for Acoustic Modeling in Speech Recognition [8]—helped articulate the shift toward deep, scalable models and representation learning that foreshadowed self-supervised ASR techniques.