Automatic Speech Attribute Detection of Arabic Language (original) (raw)
Recently, the speech attribute features caught the interest of the speech processing society and successfully employed in a wide variety of applications. In this paper we introduce the first intensive study of speech attribute detection in Arabic language. For each speech attribute, namely the manners and places of articulation, a binary Deep Neural Network (DNN) classifier is trained to recognize the existence or absence of the attribute. The DNN consists of multiple fully connected hidden layers and a two-way output softmax layer. The DNN is fed by mel-scale filter bank features extracted from the speech signal. We further adopted the dropout regularization technique to alleviate the classifier overfitting. The system tested on a speech corpus of 90 hours collected from Quranic Arabic reciters. The results show that the speech attribute detectors achieved classification accuracies ranging from 76% to 95%.
Loading Preview
Sorry, preview is currently unavailable. You can download the paper by clicking the button above.