The present invention discloses an echo cancellation method based on a regression factor of far-end estimation signals and error signals. Regression analysis of far-end estimation signals and near-and-far-end error signals is employed to obtain a regression factor which represents an estimation deviation condition of an adaptive filter. An updated optimal step length of the filter is calculated, and the regression factor comprises three voice activity states so as to omit far-and-near-end voice activity detection and dual-talk voice activity detection. The method provided by the invention canoverall plan and manage the key steps of the whole echo elimination algorithm through calculation of one parameter to greatly reduce the computing complexity. The calculation of the optimal step length is related to the data of the far end and the near end to reduce the voice distortion of the near end.