The present invention belongs to the field of voice-frequency signal processing technology, it is characterized by that it successively contains the following steps: extracting time-domain transient characteristics in signal subband interior, and extracting interframe spectrum related characteristics; utilizing time-domain transient state and interframe spectrum related value to divide error frame into two stages according to the state, it contains 10 states; according to different error frame state selecting interpolation method to generate frequency spectrum coefficient and frequency domain phase of error frame, calculating time-domain signal and front and rear edge smoothing treatment.