Automatic speech recognition (ASR) was first proposed by Bell Laboratories in 1952 through the development of the "Audrey" system, and the foundational results were published in a paper in the Proceedings of the Acoustical Society of America. Automatic Recognition of Spoken Digits.

In 1975, James K. Baker, a scholar at Carnegie Mellon University (CMU), published his doctoral dissertation. Stochastic Modeling as a Means of Automatic Speech RecognitionIt was the first time that a Hidden Markov Model (HMM) was introduced, laying the foundation for a probabilistic statistical paradigm for large-vocabulary continuous speech recognition.

In November 2012, researchers from the University of Toronto, Microsoft, Google, and IBM jointly published a landmark paper. Deep Neural Networks for Acoustic Modeling in Speech RecognitionThis paper formally established the modern ASR technology paradigm based on deep neural networks (DNNs) and was published in IEEE Signal Processing Magazine.

This technology is a framework for converting human spoken language into written text, aiming to address the "interaction gap" problem caused by machines' inability to interpret acoustic signals when facing natural human conversations. The system processes audio signals containing human speech, comprehensively utilizing acoustic models, language models, and deep neural networks to accurately identify phonemes, words, and sentences in the audio input, and then transcribes them into standardized text format. Research results show that the evolution from traditional Hidden Markov Models (HMMs) probabilistic modeling to modern Deep Neural Networks (DNN-HMMs) and end-to-end architectures effectively breaks through the performance bottlenecks of early template matching and Gaussian Mixture Models (GMMs), significantly enhancing the system's ability to interpret complex accents and environmental noise. In scenarios such as voice assistants, transcription services, voice-controlled systems, and accessibility assistance for the hearing impaired, machines achieve highly accurate and efficient speech-to-text conversion.