Donor selection for voice conversion

Voice conversion techniques enable the transformation of a source speaker´s voice to that of a target speaker´s automatically. The performance of any voice conversion algorithm depends on the source-target pair chosen. This study focuses on the problem of source speaker (donor) selection from a set of available speakers that will result in the best quality output for a specific target speaker´s voice. A voice conversion database of 20 speakers (10 male, 10 female) is collected. 180 conversions that cover all male-to-male and female-to-female voice conversion combinations are performed using a codebook mapping based method. A listening test is performed in order to determine the subjective scores for similarity of the output to the target speaker´s voice and the output quality. The results show that selecting the appropriate donor improves voice conversion performance significantly. Preliminary analysis is performed for automatic donor selection with multilayer perceptrons.