In human gait, gait asymmetry and gait speed are critical parameters that can serve as indicators of adaptability to wearable robots (e.g., robotic prostheses, exoskeletons), and are vital for controlling these robots. These parameters play a significant role in robot control and can contribute to enhancing the robot's performance. However, most current research is confined to experimental and rehabilitation environments and does not fully reflect the complexities of real-life settings. Particularly, the approach of directly attaching sensors to the body in daily life is not preferred due to privacy concerns, maintenance difficulties, and other issues. To address these problems, this study utilizes sensor data embedded in wearable robot from the affected side. Firstly, three regression models (Linear, LASSO, Ridge regression) was employed to examine the influence of independent variables on dependent variables and estimate asymmetry. Asymmetry was defined using the Symmetry Index, for which we used gait variables such as Stance time, Swing time, and Swing-Stance time ratio. Furthermore, we investigated whether the estimated values, when used as additional inputs in a learning technique combining Convolution Neural Network with Long Short-Term Memory that takes thigh angular velocity as input, could enhance the accuracy of gait speed estimation. We compared models without additional inputs, models with actual SI as additional input, and models with predicted SI as additional input. Using a knee brace, we artificially induced gait asymmetry in three conditions (normal walking, asymmetry walking with 20 degrees knee flexion restriction, asymmetry walking with 0 degree knee flexion restriction) and collected experimental data. We also used walking data from a single above-knee amputee using a robotic prosthesis to validate the estimation models. Seven healthy adults were involved in the study, attaching markers and performing level ground walking at their preferred speed. To analyze generalizability, Leave-one-subject-out cross-validation was performed. The data obtained from the experiments were used to estimate gait asymmetry by combining kinematic information of the thigh on the affected side with temporal gait variables. In the regression models, linear regression was found to be the best, and Swing-Stance time ratio was identified as the most effective gait variable for defining SI. Regarding gait speed estimation, the root mean square error values were 0.0570 m/s without additional input, 0.0515 m/s considering actual SI, and 0.0492 m/s considering estimated SI, confirming that including gait asymmetry as an additional input, i.e., SI, enhances the accuracy of gait speed estimation.