speech recognition

In PHONETICS and COMPUTATIONAL LINGUISTICS, the recognition of human SPEECH through computer analysis; also called automatic speech recognition (ASR). The term should not be confused with SPEAKER RECOGNITION, where the aim is to identify the person speaking rather than what is being said. The task involves the matching of an input acoustic signal with a vocabulary (of sounds, syllables, words, etc.) stored in the computer’s memory. A standard technique for matching individual words is to use stored waveforms (or features/parameters of waveforms) against which an input signal is matched (‘template matching’). The computer requires a period of training, in which it receives examples of spoken words provided by (single or multiple) speakers, and averages these to derive a CANONICAL waveform. The variable RATE of speech inputs needs to be taken into account, most often using the technique of ‘dynamic time warping’, in which SEGMENTS in the input signal are aligned with those in the template. The more challenging aim of ASR is to handle continuous speech. Here the computer is provided with information about typical patterns of phonetic and PHONOLOGICAL segmentation, as well as MORPHOLOGICAL and SYNTACTIC information. More advanced simulations, such as those provided by CONNECTIONIST models, are also used.