Design of the Tibetan Continuous Speech Corpus Based on Triphone

Large vocabulary continuous speech recognition system performance largely depends on the quality of speech corpus and how to select corpus is the key of corpus design. By taking Tibetan Amdo dialect in XiaHe as the research object, this paper builds continuous speech corpus based on triphone. At first, we collect text corpus with 100 thousand sentences and transform them to IPA according to pronunciation in XiaHe dialect, and then summarize the structure of triphone juncture, analyze the combination type and frequency of triphone in corpus statistically with text-processing platform in detail. At last by comprehensively considering coverage rate and sparseness of triphone and class-triphone we design the algorithm for extraction of corpus and realize automatic selection to corpus.