An improved voice conversion method based on deep belief network (DBN) was proposed, comprehensively considering the relationship between the speech frames and the effect of post-processing network. The method utilized a linear predictive analysis-synthesis model to extract the feature parameters of a speaker’s linear predictive spectrum, and the regional fusion spectral feature parameters for DBN were constructed so as to pretrain the model. Finally, an error correction network for the feature compensation of a detailed spectrum was introduced after fine-tuning. The comparison results show that, the spectral distortion of the converted speech shows the tendency of decreasing as the number of speech frames increases. Meanwhile, when the number of training speech frames was small, the spectral distortion of the proposed method was less than 50% between genders and less than 60% within genders. The experimental results showed that the spectral distortion of the proposed method was 6.5% lower than that of the traditional method. The proposed method significantly improves the naturalness and intelligibility of converted speech in view of two different subjective evaluations.