In this study, joint torques in the sagittal plane are estimated using joint angles and electromyography (EMG) signals during subjects' walk at 7 different speeds. First, a general inter-subject model is built by backpropagation neural network (BPNN) with data from 12 subjects. Then, to improve the estimation performance of the inter-subject for a new subject, sparse gaussian process (SGP) with residual estimation using input and output (RIO) kernel is used to compensate for the model as a transfer learning method. The obtained intra-subject model has superior performance with a relatively small amount of data in the training process. This article can be referenced when you have limited training data to estimate the torques on a new subject.