Observation-Model Error Compensation for Enhanced Spectral Envelope Transformation in Voice Conversion

TitleObservation-Model Error Compensation for Enhanced Spectral Envelope Transformation in Voice Conversion
Publication TypeConference Paper
Year of Publication2015
Conference NameIEEE International Workshop on Machine Learning for Signal Processing
AuthorsVillavicencio, F., Bonada J., & Hisaminato Y.
Conference Start Date17/07/2015
Conference LocationBoston, USA
AbstractThis work proposes a novel derivation of the spectral envelope transformation in Voice Conversion to alleviate degradations in the converted speech quality produced by the imposition of oversmoothed spectra. The existing mismatch between an input feature and the corresponding observation by the statistical model denotes an averaging of the features due to the model’s limited capacity to represent the feature space. The proposition is based on compensating this mismatch on the transformation applied to the input spectra. As a result, the perceived naturalness of the converted speech is enhanced. Our claim is supported by the results of objective and subjective evaluations comparing speech converted by the conventional transformation and the proposed one.
intranet