Tibetan language continuous speech recognition based on active WS-DBN

Because it is time-consuming and costly to annotate the large vocabulary Tibetan language corpus, it is not suitable to directly adopt the traditional automatic speech recognition (ASR) methods such as Hidden Markov Model (HMM), Dynamic Bayesian Networks (DBN), Artificial Neural Network (ANN). Thus, active learning can reduce annotation cost by sample selection. This paper proposed a new method to learn the Tibetan language continuous speech recognition model by combining DBN with active learning. The results of recognition experiments show that the proposed algorithm can reach the same recognition rate as traditional passive learning with few labeled training examples.