BulPhonC: Bulgarian Speech Corpus for the Development of ASR Technology

BulPhonC: Bulgarian Speech Corpus for the Development of ASR Technology Neli Hateva, Petar Mitankin, Stoyan Mihov FMI Sofia University, FMI Sofia University and IICT Bulgarian Academy of Sciences, IICT Bulgarian Academy of Sciences FMI 5 James Bourchier Blvd. 1164 Sofia Bulgaria, IICT Acad. G. Bonchev St. Block 25A 1113 Sofia Bulgaria nelly.hateva@gmail.com, petar@lml.bas.bg,stoyan@lml.bas.bg Abstract In this paper we introduce a Bulgarian speech database, which was created for the purpose of ASR technology development. The paper describes the design and the content of the speech database. We present also an empirical evaluation of the performance of a LVCSR system for Bulgarian trained on the BulPhonC data. The resource is available free for scientific usage. Keywords: Language resources, speech recognition corpus, Bulgarian speech database 1. Introduction A crucial language resource for building ASR systems is a properly created speech corpus, which is phonetically an- notated and representative. Recently a number of speech corpora for less common languages for ASR development have been created (Odriozola et al., 2014; Pinnis et al., 2014; Korvas et al., 2014). For Bulgarian ASR only few speech resources have been created – (Tanja Schultz, 2002; Mitankin et al., 2009). Both of the resources have been recorded in non soundproof environment. The number of speakers and number of utterances in those databases have been relatively small compared to ASR speech databases for other languages1. Here we present a new speech corpus created in the frame- work of the AComIn project2. The corpus has been spe- cially designed to facilitate the development of modern ASR technology. In the next section we provide a detailed description of the database. Afterwards an evaluation of the performance of the ASR system (Mitankin et al., 2009) trained on the BulPhonC data is given. Finally we present availability information and the conclusion. 2. BulPhonC: Bulgarian Speech Database 2.1. Phonetic Balancedness of the Speech Recognition Corpus The phonetic annotation of the BulPhonC corpus has used the phonetic system presented in (Mitankin et al., 2009), given on Table 1. The corpus consists of speech read from selected declar- ative and interrogative sentences. Table2 shows some ex- amples of sentences in the corpus. The sentences in the BulPhonC corpus were automatically selected from a set of 21838 sentences, which were phonetised with the Speech- Lab text-to-speech system for Bulgarian, (Andreeva et al., 2005). 1The number of speakers in the corpus in (Tanja Schultz, 2002) is 77 and each speaker read 1940 words. The number of speakers in the corpus in (Mitankin et al., 2009) is 46 and each speaker read 478 words. 2AComIn “Advanced Computing for Innovation”, grant 316087, funded by the FP7 Capacity Programme (Research Po- tential of Convergence Regions) На първото стъпало са разположени фундаментал- ните физиологически нужди - хранене и възпроиз- веждане. Спомням си, че когато веднъж в Белград исках да си купя часовник, продавачът ме запита колко часовни- ка. Кой футболист държи рекорда за най-много отбеля- зани голове в рамките на едно световно първенство? След службата в църквата свещеникът хвърля кръст във вода, а …

Content truncated to 500 words