Tibetan Multi-Dialect Speech Recognition Using Latent Regression Bayesian Network and End-To-End Mode

Abstract

We proposed a method using latent regression Bayesian network (LRBN) to extract the shared speech feature for the input of end-to-end speech recognition model. The structure of LRBN is compact and its parameter learning is fast. Compared with Convolutional Neural Network, it has a simpler and understood structure and less parameters to learn. Experimental results show that the advantage of hybrid LRBN/Bidirectional Long Short-Term Memory-Connectionist Temporal Classification architecture for Tibetan multi-dialect speech recognition, and demonstrate the LRBN is helpful to differentiate among multiple language speech sets.