Recognition Method of Wa Language Isolated Words Based on Convolutional Neural Network.

Speech recognition technology is a popular research direction in artificial intelligence, especially with the development of deep learning technology, speech recognition gradually shifts from traditional recognition methods to end-to-end recognition based on deep learning. Most of the current speech recognition models have achieved high recognition accuracy for mainstream languages, but these models are relatively complex in structure and have many model parameters, which are not suitable for recognizing isolated words in low-resource languages. Based on the deep learning approach, we use a simple and effective model to recognize isolated words in Wa language of minority languages. The encoder includes a simplified deep convolutional neural network VGG and BiLSTM, where the VGG network is used to extract depth features of the audio signal and BiLSTM is further encoded. The decoder includes two decoding methods, CTC and Attention, which can be decoded individually or jointly, which is an end-to-end speech recognition model. We use this model to conduct experiments on our Wa isolated words speech dataset, and the experimental results show that the model has a good recognition effect. The WER is below 20% whether it is decoded alone or jointly.