Voice conversion using deep neural network in super-frame feature space

Abstract

This paper presents a voice conversion technique using deep neural networks (DNNs) to map the spectral envelopes of a source speaker to that of a target speaker. Short-time spectral envelopes are represented by the linear predication cepstrum coefficients (LPCC) parameters, and neighbor frames are gathered to form super-frames. Then the powerful mapping ability of DNN which has a five-layer architecture consisting of three restricted Boltzmann machines (RBMs) was exploited to derive the spectral conversion function. A comparative study of voice conversion using a DNN model and the conventional Gaussian mixture model (GMM) is conducted. Experimental results show the speaker identification rate of conversion speech achieves 97.5% which is 0.8% higher than the performance of GMM method, and the value of average cepstrum distortion is 0.87 which is 5.4% higher than the performance of GMM method. ABX and MOS evaluations indicate that the conversion performance is better than the traditional GMM method under the parallel corpora condition.