An Automatic Blind Syllable Segmentation Model Based on Bi-directional LSTM

Most of the languages in the Sino-Tibetan language family use syllables as the basic unit of speech recognition. Recently most methods of syllable segmentation extract the features of speech signals as the basis of segmentation, and some predefined thresholds are introduced to determine the boundaries. Considering that the location of a syllable boundary is not only related to its previous speech signal sequence, but also to its later speech signals, a blind syllable segmentation method based on bi-directional recurrent neural network (BRNN) is proposed, in which the pure speech signals are used as original input. In order to alleviate the gradient vanishing problem, Long Short-Term Memory (LSTM) is used as the basic neuron unit in the BRNN model. In order to calculate the loss of the model efficiently, a linear distance factor and an auxiliary label vector are introduced. The two vectors are only related to the speech signals, and can be calculated at the same time as the training set is generated. With the help of these two vectors, a loss function based on cross-entropy is designed, and the model is trained and evaluated under the speech data with a total length of 2.8 hours coming from a Bai Language speech corpora. The experimental results show that the proposed model based on bi-directional LSTM (BiLSTM) can be effectively applied to syllable segmentation without additional threshold parameters, and the detection accuracy of syllable boundaries is close to 85% with the evaluating dataset.