Creation of Connected Word Speech Corpus for Bangla Speech Recognition Systems
Md. Farukuzzaman Khan *
Department of Computer Science and Engineering, Islamic University, Kushtia-7003, Bangladesh
M. Abdus Sobhan
Prime University, Mazar Road, Mirpur-1, Dhaka-1216, Bangladesh
*Author to whom correspondence should be addressed.
Abstract
A new speech corpus of connected Bangla words derived from newspapers text corpus BdNC01 has recorded. This has been designed for various research activities related to speaker-independent Bangla speech recognition. The database consists of speech of 100 speakers, each of them uttered 52 sentences as connected words for training database. Another 50 new speakers were employed to speak all the list of speech to construct a test database. Every utterance was repeated 5 times in various days to avoid time variation of speaker property. A total of 62 hours of recording makes the corpus largest in its type, size and application area. This paper describes the motivation for the corpus and the processes undertaken in its construction. The paper concludes with the usability of the corpus.
Keywords: Bangla, speech corpus, vocabulary, connected word, BdNC01