The Design of Continuous Speech Corpus Based on Half-Syllable Tibetan

The performance of large vocabulary continuous speech recognition system mostly depends on the quality of speech corpus, but the design of corpus ties in corpus selection. According to ancient Tibetan phonology, the continuous speech corpus based on half-syllable Tibetan is built. Each word in the ten thousands sentences of Tibetan texts is separated into initial and final. Then, the inner syllabic initial-final combinations and inter syllabic final-initial combinations are calculated. Finally, based on coverage and sparseness of half-syllabic combinations, the algorithm of corpus extraction is designed to finish the continuous speech corpus with high quality and low redundancy.