Speech Recognition Quality Estimation-based Semi-Supervised Training for Broadcast Radio Program Transcription

• Main Authors: Sing-Yue Wang, 王星月
• Other Authors: 廖元甫
• Format: Others
• Language: zh-TW
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/v665zj

碩士 === 國立臺北科技大學 === 電子工程系研究所 === 105 === It is difficult to collect enough labeled data to well train a high performance Automatic Speech Recognition. However, we can easily obtain unlimited amount of unlabeled speech data. In order to gain this advantage, the objective of this paper to use semi-supervised training to improve ASR’s performance. We using Quality Estimation to predict utterance WER. Then a subset of the unlabeled speech utterance which is predicted to have good recognition quality was added into the training data of the speech recognizer and retrain acoustic model. In experimental results, we evaluate two test data set of broadcast materials. The CER could be reduced from 25.00% to 23.61% and from 14.24% to 13.24% with QE-based data selection methods. We also retrain language model with Giga Word, the CER could be reduced from 23.61% to 23.25% and from 13.24% to 12.63%. Finally, we implement online Automatic Radio Transcriber provides speech recognition service.