Speech Attribute Recognition using Context-Dependent Modeling

Speech attributes, such as places and manners of articulation are robust against cross-speaker variation and environmental distortions. They have been used in various speech processing applications such as spoken language identification, speaker recognition and speech recognition. In this paper, we propose a method to recognize speech attributes by using a context-dependent modeling of the attributes, called bi-attributes. Experimental results on the TIMIT database show that the context-dependent modeling reduces frame classification error by 13.2% and 16.1% relatively over the context-independent modeling for manner and place classification, respectively. In addition, when fused with phone posteriors to improve phone recognition accuracy, the attribute context dependent modeling gives a 9.9% relative phone error rate reduction over the attribute context independent modeling.