AUTOMATIC SPEECH RECOGNITION: EXPERIMENTS WITH A RECOGNISER USING LINGUISTIC STATISTICS
UNIVERSITY COLL LONDON (UNITED KINGDOM)
Pagination or Media Count:
The problem of transmitting speech over communication channels with smaller information-carrying capacity than that of conventional telephone links is discussed. Bandwidth compression systems using articulatory constraints vocoders are described and this is followed by a description of devices that analyse the speech sound wave in terms of linguistic units - machines performing this task are called automatic speech recognisers. Bandwidth economy can be achieved by recognising and transmitting these linguistic units. The difficulties of automatic recognition are discussed and its processes compared with the human mechanism for speech recognition. It is suggested that, just as in human speech recognition, the performance of an automatic recogniser could be improved by using information about the statistics and the structure of the language as well as the usual acoustic cues. The design and construction of a phoneme recogniser for putting this idea to the test is described. The machine has three parts 1 the acoustic recogniser for detecting some simple phonemic cues, 2 stored knowledge about the digram frequencies of these phonemes, and 3 a device for selecting the phoneme that is most likely to occur in the light of both acoustic information and of the relevant digram frequencies. The selection is indicated on a typewriter.