English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

State-of-the-art non-autoregressive text-to-speech (TTS) models based onFastSpeech 2 can efficiently synthesise high-fidelity and natural speech. Forexpressive speech datasets however, we observe characteristic audiodistortions. We demonstrate that such artefacts are introduced to the vocoderreconstruction by over-smooth mel-spectrogram predictions, which are induced bythe choice of mean-squared-error (MSE) loss for training the mel-spectrogramdecoder. With MSE loss FastSpeech 2 is limited to learn conditional averages ofthe training distribution, which might not lie close to a natural sample if thedistribution still appears multimodal after all conditioning signals. Toalleviate this problem, we introduce TVC-GMM, a mixture model ofTrivariate-Chain Gaussian distributions, to model the residual multimodality.TVC-GMM reduces spectrogram smoothness and improves perceptual audio quality inparticular for expressive datasets as shown by both objective and subjectiveevaluation.

Towards Robust FastSpeech 2 by Modelling Residual Multimodality