The present invention relates to a system for extracting a signal unit feature using a variational inference model based on deep learning. The system comprises: a UBM learning unit; a Baum-Welch statistic calculating unit; a variational Inference model learning unit; and a feature extracting unit. According to the system proposed in the present invention, a parameter representing distribution of input may be generated by extracting a feature using a variational interference-based deep learning structure placing a random variable as a medium of input/output.
展开▼