| Summary: | 碩士 === 國立臺北科技大學 === 電子工程系研究所 === 104 === The objective of this paper to create an enhanced learning language model. Language is often identified by other similar languages (Out of set), channel environment, noise and other factors make performance decline. Part of the traditional methods have established language model LDA (Linear Discriminant Analysis) or PLDA (Probabilistic Linear Discriminant Analysis), in recent years, the common language identification systems are DNN (Deep Neural Network) architecture. On the training model method, the traditional method is to use Cross-Entropy, this paper using a weighted formula and reinforcement cost function method. lre15 task for the target language and recognize non-target language, with final validation scores game formula, do appraisals for test corpus, the lower the score, the better representatives recognition system. According to the results we have taken the best part of each architecture performance analysis to do, given LDA first official language identification system, its score of 39.033, its traditional DNN + reinforcement score of 30.1356, while the use of this paper proposed DNN + reinforcement of its score of 20.8996. The results can be found using DNN and reinforcement of this paper proposed resolution to have the best knowledge of performance.
|
|---|
