Using Speech Assessment Technique for the Validation of Taiwanese Speech Corpus
Yu-Jhe Li, Chung-Che Wang, Liangyu Chen, Jyh‐Shing Roger Jang, Ren-Yuan Lyu · 2013
This research focuses on validating a Taiwanese speech corpus by using speech recognition and assessment to automatically find the potentially problematic utterances. There are three main stages in this work: acoustic model training, speech assessment and error labeling, and performance evaluation. In the acoustic model training stage, we use the ForSD (Formosa Speech Database) ,provided by Chang Gung University (CGU), to train hidden Markov models (HMMs) as the acoustic models. Monophone, biphone (right context dependent), and triphone HMMs are tested. The recognition net is based on free syllable decoding. The best syllable accuracies of these three types of HMMs are 27.20%, 43.28%, and 45.93% respectively. In the speech assessment and error labeling stage, we use the trained triphone HMMs to assess the unvalidated parts of the dataset. And then we split the dataset as low-scored dataset, mid-scored dataset, and high-score dataset by different thresholds. For the low-scored dataset, we identify and label the possible cause of having such a lower score. We then extract features from these lower-scored utterances and train an SVM classifier to further examine if each of these low-scored utterances is to be removed. In the performance evaluation stage, we evaluate the effectiveness of finding problematic utterances by using 2 subsets of ForSD, TW01, and TW02 as the 使用語音評分技術輔助台語語料的驗證 83 training dataset and one of the following: the entire unprocessed dataset, both mid-scored and high-scored dataset, and high-scored dataset only. We use these three types of joint dataset to train and to evaluate the performance. The syllable accuracies of these three types of HMMs are 40.22%, 41.21%, 44.35% respectively. From the previous result, the disparity of syllable accuracy between the HMMs trained by unprocessed dataset and processed dataset can be 4.13%. Obviously, it proves that the processed dataset is less problematic than unprocessed dataset. We can use speech assessment automatically to find the potential problematic utterances.