City Research Online

Abstract

A hybrid computer has been programmed to recognise continuous speech. On the analogue console 10 bandpass filters, covering the audio frequency range, have been patched, and their rectified outputs may be sampled 25 times a second. The digital computer processes these filter outputs and indicates the sound being made at the sampling instant, this sound being one of a small library of twenty possible sounds.

Having acoustically analysed the speech at the sampling instant, the digital computer uses the time before the next trigger pulse to identify the sentence being spoken. It matches the symbol stream derived from the spoken sentence with symbol streams that can be produced from the grammar governing the speech. This grammar is restricted to being a regular grammar, in the formal language sense. The whole spoken sentence is matched using a dynamic programming approach, and only when the best match has been found is the sentence segmented, and the user presented with a visual representation of his speech. Since the dynamic programming method optimises the match, the segmentation too is optimal.

The recogniser is trainable in that new speakers can engage in a dialogue to give samples of their voice. One sample of each of the words in the vocabulary is chosen to be a standard, and is used in subsequent recognition.

The program has been tested by three people using sentences for commanding a simple desk calculator. Results are presented for two-, four- and six-word sentences, and suggestions made for reducing the recognition time.

Publication Type:	Thesis (Doctoral)
Subjects:	Q Science > QA Mathematics Q Science > QA Mathematics > QA75 Electronic computers. Computer science Q Science > QA Mathematics > QA76 Computer software
Departments:	School of Science & Technology > Department of Mathematics School of Science & Technology > School of Science & Technology Doctoral Theses Doctoral Theses

Export

Downloads