The invention discloses a sparse spectrum signature extraction method for a voice lie detection system. The sparse spectrum signature extraction method for a voice lie detection system includes the following steps: 1) extracting a Mel frequency spectral coefficient and a wavelet packet frequency band cepstrum coefficient of a voice signal, and integrating the Mel frequency spectral coefficient with the wavelet packet frequency band cepstrum coefficient to form a cepstrum characteristic; 2) training the cepstrum characteristic by means of a K-singular value decomposition algorithm to obtain a mixed overcomplete representation dictionary; and 3) on the mixed overcomplete representation dictionary obtained in the step 2, utilizing an orthogonal matching pursuit algorithm to perform sparse coding on the cepstrum characteristic to obtain the sparse spectrum signature. The sparse spectrum signature extraction method for a voice lie detection system can make up the deficiency of providing medium and high frequency band information by a traditional Mel frequency spectral coefficient, and can also solve the redundancy problem of a non-linear fusion parameter set to reduce the calculating complexity of a classification model.