Large Vocabulary Continuous Speech Recognition for Bulgarian

International Conference RANLP 2009 - Borovets, Bulgaria, pages 246–250 Large vocabulary continuous speech recognition for Bulgarian Petar Mitankin Institute for Parallel Processing Bulgarian Academy of Sciences petar@lml.bas.bg Stoyan Mihov Institute for Parallel Processing Bulgarian Academy of Sciences stoyan@lml.bas.bg Tinko Tinchev Faculty of Mathematics and Informatics So a University tinko@fmi.uni-so a.bg Abstract The paper presents the results of a project com- pleted by the authors for realizing a contin- uous speech recognition system for Bulgarian. The state-of-the-art speech recognition technol- ogy used in the system is discussed. Special attention is given to the problems with some speci cs of the Bulgarian language namely the large vocabulary (450000 wordforms). Some im- plementation details of the language module are given. At the end the paper provides evaluation of the accuracy of recognition. Keywords Speech recognition, language model 1 Introduction Speech is a preferable medium for enabling comfort- able and eective human-computer interface. This ex- plains the wide interest in speech technology. There are a number of implementations of computer speech synthesis systems, which deliver intelligible and nat- urally sounding voices. For Bulgarian the SpeechLab system [1] is widely used by the visually impaired peo- ple. Speech recognition is a considerably more di cult problem especially in the case of continuous speech and large vocabulary. There exist sophisticated implemen- tations for English and some other major European and Asian languages. There are no reports for a large vocabulary continuous speech recognition (LVCSR) systems for Bulgarian. This paper will present the result of a project for realizing a LVCSR system for Bulgarian, which has been executed by the authors. After some background information given in the next section we will describe the basic modules of our system. Section 3 and Section 4 will provide tech- nical details of the Bulgarian acoustic and language models correspondingly. The recognition process is discussed in Section 5. In Section 6 we present the implementation details of the language model and Sec- tion 7 shows our accuracy evaluation. The conclusion presents some general discussions and further develop- ment directions. 2 Background The problem of speech recognition can be considered as the task to nd the set of possible word sequences which sound close to the observed speech signal and are admissible according to a given language model. The recognition of isolated words is signi cantly sim- pler. This task can be solved using various techniques with modest computational eorts. The computa- tional complexity might be linear in regards of number of words in the vocabulary and no language model is required. A LVCSR system has to recognize a sequence of words from a large vocabulary without information of the word boundaries. A naive method would re- quire an exponential number of word permutations to be considered. The state-of-the-art approach to this problem is the Hidden Markov Model (HMM) framework. An HMM represent a statistical process, where the elements of a sequence of observed variables are regarded as emis- sions of the state variables in a Markov chain. The observed variables …

Content truncated to 500 words