Avtomatsko razpoznavanje govora za pregibni jezik z uporabo morfoloških jezikovnih modelov s kontekstno odvisno strukturo : doktorska disertacija

slovenian

doktorska dela

2015

Univerza v Mariboru, Fakulteta za elektrotehniko računalništvo in informatiko

In this thesis, we are focused on language modelling for automatic speech recognition in large vocabulary applications, where we are still experiencing the problem of insufficient recognition accuracy. This problem is more present in morphologically complex languages, for example Slovene. For such a system to work properly we need language models. State of the art speech recognition systems for Slovene still produce a hidher number of recognition errors that recognizers for other langauges. We see many sentences that are still understandable, but which contain syntactical errors. Often errors are present in the word endings. Therefore it seems reasonable to include morphosyntactic information into language models to reduce syntactical errors. This thesis presents the development of a speech recognition system that uses not only the usual n-gram language models for words, but also models that include part-of-speech and morphosyntactic information. We call them morphological models. We developed an algorithm that determines the best structure for such models based on perplexities for with respect to the part-of-speech categories of a words context. We say that the models have a context dependent structure. We implemented them as factored language models. Although we do not need a very large training corpus we still experience data sparsity due to the large number of possible context of a word. We therefore also developed a method for merging different context. Because of a large number of possible models structures it was also necessary to develop an algorithm for limiting the search space by gradually determining a models structure. The system is designed as a two-pass recognition algorithm, where the morphological models are used in the second pass. We developed an algorithm for a fast optimization of the systems parameters and dynamic weighting of the models scores based on a words context. We tested speech accuracy on the Slovene Broadcast news speech database. We also added a more detailed analysis of the recognition results.

07.08.2023