No description available.

<p>To achieve a robust system the variation seen for different speaking styles must be handled. An investigation of standard automatic speech recognition techniques for different speaking styles showed that lexical modelling using general-purpose variants gave small improvements, but the errors differed compared with using only one canonical pronunciation per word. Modelling the variation using the acoustic models (using context dependency and/or speaker dependent adaptation) gave a significant improvement, but the resulting performance for non-native and spontaneous speech was still far from read speech.</p><p>In this dissertation a complete data-driven approach to rule-based lexicon adaptation is presented, where the effect of the acoustic models is incorporated in the rule pruning metric. Reference and alternative transcriptions were aligned by dynamic programming, but with a data-driven method to derive the phone-to-phone substitution costs. The costs were based on the statistical co-occurrence of phones, association strength. Rules for pronunciation variation were derived from this alignment. The rules were pruned using a new metric based on acoustic log likelihood. Well trained acoustic models are capable of modelling much of the variation seen, and using the acoustic log likelihood to assess the pronunciation rules prevents the lexical modelling from adding variation already accounted for as shown for direct pronunciation variation modelling.</p><p>For the non-native task data-driven pronunciation modelling by learning pronunciation rules gave a significant performance gain. Acoustic log likelihood rule pruning performed better than rule probability pruning.</p><p>For spontaneous dictation the pronunciation variation experiments did not improve the performance. The answer to how to better model the variation for spontaneous speech seems to lie neither in the acoustical nor the lexical modelling. The main differences between read and spontaneous speech are the grammar used and disfluencies like restarts and long pauses. The language model may thus be the best starting point for more research to achieve better performance for this speaking style.</p>

To achieve a robust system the variation seen for different speaking styles must be handled. An investigation of standard automatic speech recognition techniques for different speaking styles showed that lexical modelling using general-purpose variants gave small improvements, but the errors differed compared with using only one canonical pronunciation per word. Modelling the variation using the acoustic models (using context dependency and/or speaker dependent adaptation) gave a significant improvement, but the resulting performance for non-native and spontaneous speech was still far from read speech. In this dissertation a complete data-driven approach to rule-based lexicon adaptation is presented, where the effect of the acoustic models is incorporated in the rule pruning metric. Reference and alternative transcriptions were aligned by dynamic programming, but with a data-driven method to derive the phone-to-phone substitution costs. The costs were based on the statistical co-occurrence of phones, association strength. Rules for pronunciation variation were derived from this alignment. The rules were pruned using a new metric based on acoustic log likelihood. Well trained acoustic models are capable of modelling much of the variation seen, and using the acoustic log likelihood to assess the pronunciation rules prevents the lexical modelling from adding variation already accounted for as shown for direct pronunciation variation modelling. For the non-native task data-driven pronunciation modelling by learning pronunciation rules gave a significant performance gain. Acoustic log likelihood rule pruning performed better than rule probability pruning. For spontaneous dictation the pronunciation variation experiments did not improve the performance. The answer to how to better model the variation for spontaneous speech seems to lie neither in the acoustical nor the lexical modelling. The main differences between read and spontaneous speech are the grammar used and disfluencies like restarts and long pauses. The language model may thus be the best starting point for more research to achieve better performance for this speaking style.

Analytic knowledge about the systematic variation in a language has an important place in the description of the language. Such knowledge is interesting e.g. in the language teaching domain, as a background for various types of linguistic studies, and in the development of more dynamic speech technology applications. In previous studies, the effects of single variables or relatively small groups of related variables on the pronunciation of words have been studied separately. The work described in this thesis takes a holistic perspective on pronunciation variation and focuses on a method for creating general descriptions of phone-level pronunciation in discourse context. The discourse context is defined by a large set of linguistic attributes ranging from high-level variables such as speaking style, down to the articulatory feature level. Models of phone-level pronunciation in the context of a discourse have been created for the central standard Swedish language variety. The models are represented in the form of decision trees, which are readable for both machines and humans. A data-driven approach was taken for the pronunciation modelling task, and the work involved the annotation of recorded speech with linguistic and related information. The decision tree models were induced from the annotation. An important part of the work on pronunciation modelling was also the development of a pronunciation lexicon for Swedish. In a cross-validation experiment, several sets of pronunciation models were created with access to different parts of the attributes in the annotation. The prediction accuracy of pronunciation models could be improved by 42.2% by making information from layers above the phoneme level accessible during model training. Optimal models were obtained when attributes from all layers of annotation were used. The goal for the models was to produce pronunciation representations representative for the language variety and not necessarily for the individual speakers, on whose speech the models were trained. In the cross-validation experiment, model-produced phone strings were compared to key phonetic transcripts of actual speech, and the phone error rate was defined as the share of discrepancies between the respective phone strings. Thus, the phone error rate is the sum of actual errors and discrepancies resulting from desired adaptations from a speaker-specific pronunciation to a pronunciation reflecting general traits of the language variety. The optimal models gave an average phone error rate of 8.2%. / QC 20100901