Error Analysis and Improving Speech Recognition for Latvian Language

Proceedings of Recent Advances in Natural Language Processing, pages 563–569, Hissar, Bulgaria, Sep 7–9 2015. Error Analysis and Improving Speech Recognition for Latvian language Askars Salimbajevs Tilde, Vienibas gatve 75a, Riga, Latvia askars.salimbajevs@tilde.lv Jevgenijs Strigins Tilde, Vienibas gatve 75a, Riga, Latvia jevgenijs.strigins@tilde.lv Abstract Developing a large vocabulary automatic speech recognition system is a very difficult task, due to the high variations in domain and acoustic variability. This task is even more difficult for the Latvian language, which is very rich morphologically and in which one word can have dozens of surface forms. Although there is some research on speech recognition for Latvian, Latvian ASR remains behind “big” languages such as English, German etc. In order to improve the performance of Latvian ASR, it is important to understand what errors does it make and why. In this paper, the authors analyze the most common errors of Latvian ASR. Based on this, baseline system WER is improved from 30.94% to 28.43%. 1 Introduction When developing an Automatic Speech Recognition (ASR) system it is typical to evaluate system performance by calculating quantitative measures like accuracy, F1 score, Word Error Rate (WER) etc. However, in order to improve ASR performance, it is important to understand which factors are most problematic for recognition, identify the types of errors, their main causes and how critical these errors are. This is even more important when developing ASR for a language, for which no such analysis has ever been done, because the developer might not know what problems to expect, where more effort and focus is needed and what possible solutions there are. The Latvian language is a moderately inflected language, with complex nominal and verbal morphology. Latvian also has a selection of prefixes and suffixes that can modify nouns, adjectives, adverbs and verbs. There is no definite or indefinite article in Latvian, but definiteness can be indicated by the endings of adjectives. Because of these properties, one word in Latvian can have tens or even hundreds (in the case of verbs) of surface forms. For example, a word “cat” in English has 3 surface forms: cat, cats and cat’s, but in Latvian the variation is much bigger: kaķis, kaķa, kaķim, kaķi, kaķī, kaķu, kaķiem etc. They all describe an animal – cat, but in the same time these are different surface forms that change the meaning of sentence. To the best of our knowledge there has been no research conducted on analyzing misrecognized words for Latvian LVASR. In fact, there are only a few published results on speech recognition for Latvian (Oparin et al., 2013; Darǵis, R., & Znotiņš, A., 2014; Salimbajevs & Pinnis, 2014), that report the best performance of WER 20.2%. However, there are no lack of efforts on error analysis for “bigger” languages (Goldwater et al., 2010; Vasilescu et al, 2012). In many cases factors discovered in these works also apply to “smaller” languages. For example there are results for English (Fosler-Lussier & Morgan, 1999) and Japanese (Shinozaki & Furui, 2001) that show that …

Content truncated to 500 words