Changing a talker's apparent age or gender in generated speech while keeping the rest of their vocal identity recognizable, the audio counterpart of editing attributes on a generated face. No parallel data exists for it — no recording shows one person at two ages — so the mapping cannot be supervised by pairing. The approach here splits the problem in two. A GAN-trained speaker representation variational autoencoder factors the speaker vector from a self-supervised speaker-recognition model into age, gender, and identity embeddings, supervising the attribute branches by classification and the identity branch by a contrastive loss, and its decoder synthesizes speaker vectors for people who do not exist from attribute labels alone. Instead of forcing attribute and identity apart with gradient reversal or mutual-information minimization, a cyclic consistency step freezes the encoder and pushes fake-label decoder outputs back into the real speaker-vector domain. A two-stage voice conversion model then generates attribute-dependent average acoustic features from phonetic posteriorgrams and pitch, and refines them into one speaker's voice with a transformer ODE trained by flow matching. The task is governed by a three-way trade-off between attribute-label consistency, speaker-identity consistency, and speech quality.
Speaker attribute control is the task of changing a talker's apparent age or gender in synthesized speech while keeping the rest of their vocal identity recognizable, the audio counterpart of editing…