基于注意力机制的大同方言语音翻译模型研究

Research on Attention-Based Speech Translation Model of Datong Dialect

大同方言与普通话在语法、发音等方面有着较大的差异,且语料资源较少,制定一套系统的大同方言语言标准异常困难.另外,从保护地方方言和增进语言交流的角度出发,以大同方言的语音翻译研究为例,旨在进一步研究方言十分丰富的山西地方方言的语音翻译,为保护山西方言这一非物质文化遗产,增进山西与全国乃至全世界的无障碍语言交流,提供一定的模型与技术支撑.针对这些问题,提出了一个类似于基于注意力机制的端到端语音识别的语音翻译模型,用于将大同方言语音翻译为普通话文本,并省去了制定语言标准的麻烦.实验表明,相对于传统的级联模型,在基于同样的实验数据的情况下,端到端模型的输出结果的双语评估替换分数有了一定的提高.

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.