Techniques and devices for automatic speech recognition

For some decades the possibility of automatic speech recognition has intrigued many speech researchers as well as the general public. This dissertation, which contains ten different speech studies and a summary, is mostly about "traditional speech recognition" which means the conversion of the acoustic speech signal to phonetic features, phonemes and words. This is opposed to 'kpeech understanding, "in which the system is required to make adequate responses to the 'heaning " of the input speech and that demands linguistic components such as syntax, semantics and pragma tics. The parameterisation of the speech signal is handled by the acoustic front end of a recognition system. The standard technique has been to convert the speech sipal to spectral information, for instance, filfer bank amplitudes. Frequently, cepstral coefficients are derived instead of amplitude parameters. These traditional techniques for speech parameterisation are basically static descr~~tions. In one study it is shown that the addition of spectral change parameters (differentia fed cepstrum coefficients) to regular cepstrum coefficients improves recognition performance. Another study attempted the use of parameters inspired by models of human auditory processing. However, there was a negative correlation befween the complexiq of the models used and the recognition performance, although some of the more simplistic models performed well. The possibility of using synthetic speech to build templates for a word recognition system has also been examined with some encouraging results. This technique has some attractive qualities, since it could dramatically reduce the need for training of a recogniser. Some papers are oriented more towards the applications of word recognition. One is a feasibility study of voice controlled dialling in an intercom system and one is about the construction of a portable hardware recogniser. Another deals with the reduction of the effeck of background noise, especially when using speech recognition for voice controlled dialling in a cars. The method used is shown to improve the noise immunify of the system. There are furthermore some papers in which artificial neural networks are used for phoneme and word recognition. It is shown that including coarticulation effects befween phonemes is a crucial factor and that in particular recurrent neural networks seem to be useful for this. Some experiments have been carried out to evaluate whether articulatorily based phonetic features are a better basis than phonemes for speaker independent word recognition. However, it seems that this is not the case, alihough the results are not fully conclusive. The last two studies deal with some linguistic aspects of speech technology. The first is a comparison of the ')honetic" word space of five European languages, especially in the context of speech recognition with broad, phonetic classes, and the results indicate that the lexical space is very unevenly exploited. The second compares word class assipment by explicit symbolic rules to an artificial neural network trained for the same task. The network-based method is shown toperform somewhat better.