A method for restoring loss feature for a strong voice recognition according to the present invention comprises the steps of: forming one frame by the observation data in the form of a spectrum vector, receiving the observation sequence made by a plurality of frames being arranged in an orderly manner as time passes, outputting confidence components as they are based on the information about a state index for the current frame, and outputting non-confidence components by minimizing the same; and further comprising the step of: estimating the values of final non-confidence components by adding them after multiplying posterior probability of every state if the non-confidence components are smaller than the values of the non-confidence components of the observation data, the values of confidence components are given to every frame, and the state index of the current frame is determined. The present invention is designed to provide a method and an apparatus for restoring loss feature for a strong voice recognition by using frequency and time dependency of voice through a hidden Markov model.
展开▼