語碼轉換文本資料增強於中英混合語音辨識;Code-Switching Data Augmentation For Chinese-English Speech Recognition

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/88309

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/88309

Title:	語碼轉換文本資料增強於中英混合語音辨識;Code-Switching Data Augmentation For Chinese-English Speech Recognition
Authors:	王騰輝;Tenghui, Wang
Contributors:	資訊工程學系
Keywords:	語碼轉換;語音辨識;基於變換器的雙向編碼器表示技術;自然語言處理;Code-Switching;Speech Recognition;BERT;NLP
Date:	2022-01-25
Issue Date:	2022-07-13 22:46:11 (UTC+8)
Publisher:	國立中央大學
Abstract:	語碼轉換是一種常見的語言表達模式，即對話中出現兩種以上的語言交錯使用。中文和英語作為世界上的主流語言，中英混合的語言表達方式更是日常的對話交流中不可或缺的，然而不同的族群、環境、說話習慣都會影響語碼轉換的組成特性。目前語音辨識研究中對此類語碼轉換文本的訓練語料依舊稀缺，因此本論文將利用類神經網路訓練出一個生成器生成語碼轉換文本，並作為kaldi toolkit語言模型的資料增強語料以達到改善中英混合辨識率的目的。本論文使用的方法是基於SEAME語料庫中的中英文本，訓練 BERT-BiLSTM-CRF模型以得知語碼轉換位置，再將純中文文本應該被進行語碼轉化的位置翻譯成英文，藉此產生符合此語料庫特性的句子。 ;Code-switching is a common mode of language expression, that is, the interleaved use of more than two languages in a conversation. Chinese and English are the mainstream languages in the world, and Chinese-English mixed language expressions are indispensable in daily dialogue and communication. However, different ethnic groups, environments, and speaking habits will affect the composition of code-switching characteristics. At present, the training corpus of such code-switched texts in speech recognition research is still scarce. Therefore, this paper will use a neural network to train a generator to generate code-switched texts, and use them as the training data of the kaldi toolkit language model .To achieve the purpose of better accuracy of Chinese-English.speech recognition. The method used in this paper is based on the Chinese-English Code-switching corpus in the SEAME dataset. We training the BERT-BiLSTM-CRF model to know the Code-switching position, and then translate the Chinese words into English words, thereby generating Sentences that match the characteristics of this corpus.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	55	View/Open

社群 sharing

Loading...