Đang chuẩn bị nút TẢI XUỐNG, xin hãy chờ
Tải xuống
In this paper, an approach of automatically correcting recognition errors of repeated words is proposed by exploiting recognition results of preceding utterances. During the error correction, the words that might appear again in the following utterances are collected from the recognition results of preceding utterances. For each utterance, there are four steps involved in the correction: 1) initial recognition. | Journal of Automation and Control Engineering Vol. 4, No. 2, April 2016 Automatic Error Correction for Repeated Words in Mandarin Speech Recognition Xiangdong Wang, Hong Liu, and Yueliang Qian Institute of Computing Technology, Chinese Academy of Sciences, Beijing Key Laboratory of Mobile Computing and Pervasive Device, Beijing, China Email: {xdwang, hliu, ylqian}@ict.ac.cn Xinhui Li Tencent Inc., China Email: hiccupli@tencent.com repeated words. For a repeated word, since the context and pronunciation varies much in different appearances, the recognition results of the same repeated word might be different in different utterances. Even though it is recognized correctly at its first appearance, it still might be recognized falsely in the following utterances. In the existing error correction techniques [1]-[4], each time the repeated word is recognized falsely in the task, the user will be required to correct the recognition error. If there are a lot of words which are difficult to be recognized correctly and appear repeatedly in the task, such as professional terms and name entities, correcting these errors using the existing error correction techniques will cost the user a lot of time. In order to improve the efficiency of error correction, we propose an approach which automatically corrects recognition errors of repeated words by using their prior recognition results and correction results. In this method, when the word first appeared in the speech, its recognition result and correction result will be saved in a word template database. For each utterance, there are four steps involved in the correction: 1) recognizing the speech with a well-trained recognizer, 2) detecting repeated words by computing the phonetic similarity between recognition result and word templates, 3) correcting recognition errors of repeated words automatically, 4) extracting new words using the text after manual error correction or extract new words without any additional correction, .
