A mechanomyogram is a visualization of the mechanical signal from the surface of a muscle when the muscle is contracted. The setup of the mechanomyography (MMG) measurement is simpler than the setup for surface electromyography (sEMG) measurement and is less affected by sweating. However, torque estimation based on a mechanomyogram involves significant noise, which is an important issue. Therefore, we propose a regression analysis method to estimate the torque of the knee joint during voluntary movement based on the MMG signal. The proposed method differs from conventional methods because it integrates the MMG sensor responses at four locations: anterior, posterior, and medial/lateral just above the main operating muscle. This method focuses on the acceleration response characteristics, which change slightly depending on the location of the MMG sensor. Support vector regression was performed on the root mean square (RMS) of the MMG signals, which were processed by a low-pass filter. Two-channel estimation with an increased number of MMG sensors for the leading and antagonist muscles improved the conventional method, and four-channel estimation with medial and lateral sensors further improved the performance. These results show that the estimation performance of the proposed method does not significantly differ from that of the surface electromyogram.