| Summary: | We address a nonparallel data-driven many-to-many speech modeling and multimodal style conversion method. In this work, we train a speech conversion model for multiple domains rather than a specific source and target domain pair, and we generate diverse output speech signals from a given source domain speech by transferring some speech style-related characteristics while preserving its linguistic content information. The proposed method comprises a variational autoencoder (VAE)-based many-to-many speech conversion network with a Wasserstein generative adversarial network (WGAN) and a skip-connected autoencoder-based self-supervised learning network. The proposed conversion network trains the models by decomposing the spectral features of the input speech signal into a content factor that represents domain-invariant information and a style factor that represents domain-related information to automatically estimate the various speech styles of each domain, and the network converts the input speech signal to another domain using the computed content factor with the target style factor we want to change. Diverse and multimodal outputs can be generated by sampling different style factors. We also train models in a stable manner and improve the quality of generated outputs by sharing the discriminator of the VAE-based speech conversion network and that of the self-supervised learning network. We apply the proposed method to speaker conversion and perform the perceptual evaluations. Experimental results revealed that the proposed method obtained high accuracy of converted spectra, significantly improved the sound quality and speaker similarity of the converted speech, and contributed to stable model training.
|
|---|
