Voice Conversion Based on Unified Dictionary with Clustered Features Between Non-parallel Corpus

Non-negative matrix factorization (NMF) has been widely applied to exemplar-based voice conversion(VC) recently. It differs noise robustness and naturalness of the converted voice, compared with conventional statistical Gaussian mixture model-based VC. However, parallel training data from source and target speakers are required so it can not realize the arbitrary speakers' voice conversion, especially when the corpus of target speakers is inadequate. In this paper, we present a novel algorithm by clustering the spectral features in high dimensions to construct the unified dictionary and introduce a mapping matrix between source and target sparse coefficients. Experimental results demonstrate that the value of average cepstral distortion is 0.833 which is about 4.3% lower than the performance of conventional NMF based method. Subjective evaluations such as ABX and MOS are also discussed. It indicates that the speech quality in our study is quite better than conventional NMF. The target speaker's spectra are even unnecessary to be included in the training set.