In order to fully consider the inter-frame correlation in voice conversion, a convolutive nonnegative matrix factorization based voice conversion method is proposed. The personal characteristics and inter-frame correlation in voice can be well preserved in the time-frequency bases obtained from convolutive nonnegative matrix factorization. With this feature, during the training phase of voice conversion, the matching time-frequency bases of source and target speakers can be extracted from training data through convolutive nonnegative matrix factorization. Then in the conversion phase, the voice of source speaker is converted through time-frequency bases substitution. Compared to traditional methods, the inter-frame correlation in voice can be better preserved and converted in the proposed method. Experimental results using objective and subjective evaluations show that the proposed method outperforms the Gaussian Mixture Model and State Space Model based methods in the view of both speech quality and conversion similarity to the target speech.