2017/01/30 by Shrikant Malviya, Rohit Kumar Mishra, Malviya, Shrikant +3 · 1 citation
Computer Science · #Speech Recognition and Synthesis #Text and Document Classification Technologies #Natural Language Processing Techniques
paper · pdf · doi:10.48550/arxiv.1701.08655
Automatic speech recognition (ASR) and Text to speech (TTS) are two prominent\narea of research in human computer interaction nowadays. A set of phonetically\nrich sentences is in a matter of importance in order to develop these two\ninteractive modules of HCI. Essentially, the set of phonetically rich sentences\nhas to cover all possible phone units distributed uniformly. Selecting such a\nset from a big corpus with maintaining phonetic characteristic based similarity\nis still a challenging problem. The major objective of this paper is to devise\na criteria in order to select a set of sentences encompassing all phonetic\naspects of a corpus with size as minimum as possible. First, this paper\npresents a statistical analysis of Hindi phonetics by observing the structural\ncharacteristics. Further a two stage algorithm is proposed to extract\nphonetically rich sentences with a high variety of triphones from the EMILLE\nHindi corpus. The algorithm consists of a distance measuring criteria to select\na sentence in order to improve the triphone distribution. Moreover, a special\npreprocessing method is proposed to score each triphone in terms of inverse\nprobability in order to fasten the algorithm. The results show that the\napproach efficiently build uniformly distributed phonetically-rich corpus with\noptimum number of sentences.\n