International audience

International audience

Training of multi-speaker text-to-speech (TTS) systems relies on curated datasets based on high-quality recordings or audiobooks. Such datasets often lack speaker diversity and are expensive to collect. As an alternative, recent studies have leveraged the availability of large, crowdsourced automatic speech recognition (ASR) datasets. A major problem with such datasets is the presence of noisy and/or distorted samples, which degrade TTS quality. In this paper, we propose to automatically select high-quality training samples using a non-intrusive mean opinion score (MOS) estimator, WV-MOS. We show the viability of this approach for training a multi-speaker GlowTTS model on the Common Voice English dataset. Our approach improves the overall quality of generated utterances by 1.26 MOS point with respect to training on all the samples and by 0.35 MOS point with respect to training on the LibriTTS dataset. This opens the door to au-tomatic TTS dataset curation for a wider range of languages.

International audience

In recent years, the performance of speech synthesis systems has been improved thanks to deep learning-based models, but generating expressive audiovisual speech is still an open issue. The variational auto-encoders (VAE)s are recently proposed to learn latent representations of data. In this paper, we present a system for expressive text-to-audiovisual speech synthesis that learns a latent embedding space of emotions using a conditional generative model based on the variational auto-encoder framework. When conditioned on textual input, the VAE is able to learn an embedded representation that captures emotion characteristics from the signal, while being invariant to the phonetic content of the utterances. We applied this method in an unsuper-vised manner to generate duration, acoustic and visual features of speech. This conditional variational auto-encoder (CVAE) has been used to blend emotions together. This model was able to generate nuances of a given emotion or to generate new emotions that do not exist in our database. We conducted three perceptive experiments to evaluate our findings.

International audience

(accepted)

International audience

International audience

This paper presents a Non-Uniform Units selection-based Text-To-Speech synthesizer. Nowadays, systems use prosodic models that do not allow the prosody to vary as far as we should hope, involving a listening comfort degradation. Our system has the advantage to avoid the using of prosodic model. Speech units selection builds its features set exclusively from the linguistic information generated by the natural language analysis. We also present an original method to automatically weight these features. Therefore, selected units are not restricted by a predetermined prosody. With only using linguistic features, we obtain a various prosody and the units concatenation is performed without resort to heavy signal processing

Colloque avec actes et comité de lecture. internationale.

This paper presents a speech signal transformation which slows down speech signals selectively and enhances some important acoustic cues. This transformation can be used not only for hearing aids but also for second language acquisition by facilitating oral comprehension. Selective slowing down relies on the use of the TD-PSOLA synthesis method. An automatic pitch marking algorithm was designed to apply this method automatically. The strategy used to control slowing down exploits a spectral variation function which locates rapid spectral changes. The enhancement simply consists of amplifying stop bursts and unvoiced fricatives. These acoustic cues are detected automatically through the examination of energy criteria. This approach was evaluated in the context of second language acquisition, more precisely by evaluating improvements in oral comprehension. Transformations triggered properly, i.e. the signal regions modified are those which were expected to be modified. Experiments show that the oral comprehension is improved.