Attention Mechanism; Datong Dialect; End-to-end Model; Language Communication; Speech Translation;

https://www.scopus.com/inward/record.uri?eid=2-s2.0-85085065051&doi=10.3969%2fj.issn.1673-3193.2020.03.009&partnerID=40&md5=4809ec74fb7463e2d6ba93b38972c8b9

10.3969/j.issn.1673-3193.2020.03.009

Datong dialect and Mandarin had great differences in grammar and pronunciation, and because of a small amount of corpus resources, it was extremely difficult to formulate a systematic Datong dialect language standard. In addition, from the perspective of protecting local dialects and improving language communication, this paper took the speech translation of Datong dialect as an example to further study the speech translation of Shanxi local dialects with rich dialects, in order to protect the Shanxi dialect, which was one of the intangible cultural heritage, and enhanced the barrier-free language communication between Shanxi and the whole country as well as the whole world, it provided certain models and technical support. Aiming at these problems, a speech translation model similar to attention-based end-to-end speech recognition was proposed to translate the Datong dialect speech into Mandarin text, and saved the trouble of setting language standards. Experiments show that compared to the traditional cascading model, the BLEU score of the output of the result of the end-to-end model is improved based on the same experimental data.