Hill, D.R.


Automatic speech recognition: a problem for machine intelligence

Classics

The outline is drawn of a hypothetical machine to recognise speech, comprising a basic recogniser working on short segments of acoustic waveform only, on to which may be added further structures to use knowledge of speaker characteristics, speech statistics, syntax rules, and semantics, in order to improve the recognition performance. Suppose one tried to implement a recogniser by telling the machine to store every new pattern it encountered together with a label telling it what word or words the pattern represented, with the intention of recognising an arbitrarily large vocabulary for an arbitrarily large proportion of the total population of speakers. The fifth section will describe briefly some work which is being carried out at Standard Telecommunication Laboratories towards implementing a real machine, and the final section will contain conclusions. If, as is highly probable for ASR, the speech is transmitted through a telephone link the problems of noise and distortion can be quite severe and include noises due to handling the handset, clicks and hisses in the speech band, limitation of the bandwidth to the range between 300 and 3400 cps, and pre-emphasis of the signal.