Voice Conversion, Voice Morphing, and Speech Style Modification

A speech sample can be morphed from one voice to another by computing a sufficiently detailed time-frequency parameterization of the signal, optionally decomposing it into speaker-dependent and speaker-independent signal components, editing the parameterization or signal components, and then generating a speech signal based on the edited parameterization. Fine-grained control of time-frequency parameters has found applications in the scientific study of speech perception and in artistic applications such as voiceover and music. Coarse-grained control of speaker identity, emotion, and second-language accent have been applied to various tasks, including the training of speech recognizers and speaker anonymization.