Abstract: Heart sound segmentation is a prerequisite for accurate heart sound classification. Aimed at heart sound segmentation, an algorithm based on Bi-LSTM and state constraints is proposed. The optimal parameters of Bi-LSTM network were determined by grid method, and the heart sound state recognition model was trained. The duration of the heart sound state predicted by Bi-LSTM was counted, and the autocorrelation parameters were calculated. The autocorrelation parameters and heart sound inherent state transition rules were used to constrain the predicted heart sound state. Using the five-fold cross-validation method to test on the PhysioNet/CinC2016 data set, the algorithm has better overall performance than similar algorithms.