In a speech decoding method and apparatus, an adaptive code vector is obtained from an adaptive codebook, and a time series vector is obtained from an excitation codebook. Gains of the adaptive code vector and an excitation code vector are respectively decoded from a gain code. The gain of the adaptive code vector is classified into a first gain corresponding to a first noise level or a second gain corresponding to a second noise level. A value is determined based on the classifying results, and a mathematical operation is performed on the time series vector and the determined value. The adaptive code vector and the time series vector are weighted by the decoded gains, and an excitation signal is obtained by adding the weighted adaptive code vector and the weighted time series vector. A speech is synthesized using the excitation signal and a decoded linear prediction parameter.