 This paper proposes a novel labeling method for human gate phase estimation which can be used to train a deep learning model. The proposed labeling method uses a piecewise linear label to capture the ground truth of the user's gate phase at different walking speeds. When applied to a LSTM model, the proposed labeling method improves the accuracy of the estimated gate phase, especially during the mid-stance phase. Additionally, it maintains high accuracy in detecting the heel strike and toe off. This article was authored by Wulim Hong, Jin Won Lee, and Pil Wen He.