A robust method and apparatus to detect voice activity based on the power level of an audio frame. The method may include performing primary active/non-active voice period determination of an input audio frame according to a power level of the audio frame, extracting a noise power prediction value and a signal power prediction value by referring to power levels of current and previous audio frames according to a primary active/non-active voice period determination value, and performing secondary active/non-active voice period determination for the input audio frame by comparing the extracted signal power prediction value with the extracted noise power prediction value.