Effective and direct control of neural TTS prosody by removing interactions between different attributes

End-to-end TTS advancement has shown that synthesized speech prosody can be controlled by conditioning the decoder with speech prosody attribute labels. However, to annotate quantitatively the prosody patterns of a large set of training data is both time consuming and expensive. To use unannotated data, variational autoencoder (VAE) has been proposed to model individual prosody attribute as a random variable in the latent space. The VAE is an unsupervised approach and the corresponding latent variables are in general correlated with each other. For more effective and direct control of speech prosody along each attribute dimension, it is highly desirable to disentangle the correlated latent variables. Additionally, being able to interpret the disentangled attributes as speech perceptual cues is useful for designing more efficient prosody control of TTS. In this paper, we propose two attribute separation schemes: (1) using 3 separate VAEs to model the real-valued, different prosodic features, i.e., F₀, energy and duration; (2) minimizing mutual information between different prosody attributes to remove their mutual correlations, for facilitating more direct prosody control. Experimental results confirm that the two proposed schemes can indeed make individual prosody attributes more interpretable and direct TTS prosody control more effective. The improvements are measured objectively by F₀ Frame Error (FFE) and subjectively with MOS and A/B comparison listening tests, respectively. The scatter diagrams of t-SNE also demonstrate the correlations between prosody attributes, which are well disentangled by minimizing their mutual information. Synthesized TTS samples can be found at https://xiaochunan.github.io/prosody/index.html.