Exploring the multidimensional representation of individual speech acoustic parameters extracted by deep unsupervised models
Résumé
Understanding latent representations of speech by unsupervised models enables powerful signal analysis, transformation, and generation. Numerous studies have identified directions of variation of acoustic features such as fundamental frequency or formants in unsupervised models latent spaces, but it is yet not well understood why the variation of such one-dimensional features is often explained by multiple latent dimensions. This paper proposes a methodology for interpreting these dimensions, in the latent space of a variational autoencoder trained on a multi-speaker database.
Origine : Fichiers éditeurs autorisés sur une archive ouverte