White Rose University Consortium logo
University of Leeds logo University of Sheffield logo York University logo

Context-aware speech synthesis: A human-inspired model for monitoring and adapting synthetic speech

Nicolao, Mauro (2019) Context-aware speech synthesis: A human-inspired model for monitoring and adapting synthetic speech. PhD thesis, University of Sheffield.

[img]
Preview
Text
MauroNicolao_phdThesis.pdf
Available under License Creative Commons Attribution-Noncommercial-No Derivative Works 2.0 UK: England & Wales.

Download (10Mb) | Preview

Abstract

The aim of this PhD thesis is to illustrate the development a computational model for speech synthesis, which mimics the behaviour of human speaker when they adapt their production to their communicative conditions. The PhD project was motivated by the observed differences between state-of-the- art synthesiser’s speech and human production. In particular, synthesiser outcome does not exhibit any adaptation to communicative context such as environmental disturbances, listener’s needs, or speech content meanings, as the human speech does. No evaluation is performed by standard synthesisers to check whether their production is suitable for the communication requirements. Inspired by Lindblom's Hyper and Hypo articulation theory (H&H) theory of speech production, the computational model of Hyper and Hypo articulation theory (C2H) is proposed. This novel computational model for automatic speech production is designed to monitor its outcome and to be able to control the effort involved in the synthetic speech generation. Speech transformations are based on the hypothesis that low-effort attractors for a human speech production system can be identified. Such acoustic configurations are close to minimum possible effort that a speaker can make in speech production. The interpolation/extrapolation along the key dimension of hypo/hyper-articulation can be motivated by energetic considerations of phonetic contrast. The complete reactive speech synthesis is enabled by adding a negative perception feedback loop to the speech production chain in order to constantly assess the communicative effectiveness of the proposed adaptation. The distance to the original communicative intents is the control signal that drives the speech transformations. A hidden Markov model (HMM)-based speech synthesiser along with the continuous adaptation of its statistical models is used to implement the C2H model. A standard version of the synthesis software does not allow for transformations of speech during the parameter generation. Therefore, the generation algorithm of one the most well-known speech synthesis frameworks, HMM/DNN-based speech synthesis framework (HTS), is modified. The short-time implementation of speech intelligibility index (SII), named extended speech intelligibility index (eSII), is also chosen as the main perception measure in the feedback loop to control the transformation. The effectiveness of the proposed model is tested by performing acoustic analysis, objective, and subjective evaluations. A key assessment is to measure the control of the speech clarity in noisy condition, and the similarities between the emerging modifications and human behaviour. Two objective scoring methods are used to assess the speech intelligibility of the implemented system: the speech intelligibility index (SII) and the index based upon the Dau measure (Dau). Results indicate that the intelligibility of C2H-generated speech can be continuously controlled. The effectiveness of reactive speech synthesis and of the phonetic contrast motivated transforms is confirmed by the acoustic and objective results. More precisely, in the maximum-strength hyper-articulation transformations, the improvement with respect to non-adapted speech is above 10% for all intelligibility indices and tested noise conditions.

Item Type: Thesis (PhD)
Academic Units: The University of Sheffield > Faculty of Engineering (Sheffield) > Computer Science (Sheffield)
The University of Sheffield > Faculty of Science (Sheffield) > Computer Science (Sheffield)
Depositing User: Mauro Nicolao
Date Deposited: 19 Aug 2019 08:13
Last Modified: 19 Aug 2019 08:13
URI: http://etheses.whiterose.ac.uk/id/eprint/24678

You do not need to contact us to get a copy of this thesis. Please use the 'Download' link(s) above to get a copy.
You can contact us about this thesis. If you need to make a general enquiry, please see the Contact us page.

Actions (repository staff only: login required)