Modelling the effects of speech rate variation for automatic speech recognition

In automatic speech recognition it is a widely observed phenomenon that variations in speech rate cause severe degradations of the speech recognition performance. This is due to the fact that standard stochastic based speech recognition systems specialise on average speech rate. Although many approaches to modelling speech rate variation have been made, an integrated approach in a substantial system still has be to developed. General approaches to rate modelling are based on rate dependent models which are trained with rate specific subsets of the training data. During decoding a signal based rate estimation is performed according to which the set of rate dependent models is selected. While such approaches are able to reduce the word error rate significantly, they suffer from shortcomings such as the reduction of training data and the expensive training and decoding procedure. However, phonetic investigations show that there is a systematic relationship between speech rate and the acoustic characteristics of speech. In fast speech a tendency of reduction can be observed which can be described in more detail as a centralisation effect and an increase in coarticulation. Centralisation means that the formant frequencies of vowels tend to shift towards the vowel space center while increased coarticulation denotes the tendency of the spectral features of a vowel to shift towards those of its phonemic neighbour. The goal of this work is to investigate the possibility to incorporate the knowledge of the systematic nature of the influence of speech rate variation on the acoustic features in speech rate modelling. In an acoustic-phonetic analysis of a large corpus of spontaneous speech it was shown that an increased degree of the two effects of centralisation and coarticulation can be found in fast speech. Several measures for these effects were developed and used in speech recognition experiments with rate dependent models. A thorough investigation of rate dependent models showed that with duration and coarticulation based measures significant increases of the performance could be achieved. It was shown that by the use of different measures the models were adapted either to centralisation or coarticulation. Further experiments showed that by a more detailed modelling with more rate classes a further improvement can be achieved. It was also observed that a general basis for the models is needed before rate adaptation can be performed. In a comparison to other sources of acoustic variation it was shown that the effects of speech rate are as severe as those of speaker variation and environmental noise. All these results show that for a more substantial system that models rate variations accurately it is necessary to focus on both, durational and spectral effects. The systematic nature of the effects indicates that a continuous modelling is possible.