This thesis relates to text-to-speech synthesis and deals more particularly with the corpus based approach. In the last few years, this approach based on the concatenation of acoustic segments contained in large databases has become increasingly popular. Indeed, selecting units which best fit the text to be synthesized leads to a synthesised signal whose naturalness can be rather well preserved. The quality of the synthesized speech obtained by corpus-based methods is closely related on the one hand to the corpus used for synthesis and on the other hand to the unit selection algorithm. In spite of the notable increase of quality reached with this technology, corpus-based speech synthesis is not able to guarantee a synthesised speech whose quality is constant on an entire utterance. This is mainly due to the lack of acoustic control of the existing corpus-based speech synthesis systems. The main objective of this thesis is therefore to introduce a mechanism allowing a better acoustical control during synthesis. The proposed method uses statistical approaches to generate a smooth acoustic target from which the sequence of synthesis units will be selected. This target is deduced from acoustic models, namely context dependent senone models, estimated during a training phase. Initially, we propose an algorithm of selection based only on this acoustic target. Then, the proposed selection method is modified so as to better control the information of fundamental frequency. This unit selection module is also combined with a pre-selection module so as to drastically reduce the computational load. Formal listening tests show that the proposed method leads to a significant reduction in acoustic discontinuities during the concatenation. The proposed method is also applied to acoustic database reduction and enables a compression of about 60% of the acoustic database without perceptible decrease of the speech quality. Liste des tableaux 5.3 Moyennes etécarts types de la distorsion spectrale aux points de concaténations de la méthode proposée avec les différentes bases. .

In the domain of content delivery over Internet, each of the Client/Server and P2P communication modes has its pros and cons. In this scope, hybrid network architectures have been recently proposed as a relevant solution. In this paper we propose a new hybrid architecture that is called P2PWeb, between the centralized client/server and the non-centralized P2P architectures for content delivery. The main objective of this proposal is to reduce the load over the server in order to provide a better Quality of Service (QoS) for the end-users. A new P2PWeb communication protocol has been implemented and deployed to reach the objective. The experimentation results and the performance evaluations that we have made show the efficiency of the proposed system in terms of QoS evaluations.

No abstract

No abstract

A hybrid CDN/Viewer-to-Viewer (V2V) architecture is an attractive solution for HTTP (HLS) and MPEG-DASHbased live streaming providers. It combines a traditional CDN with a V2V overlay for exchanging video fragments, reducing the cost of the CDN while maintaining the quality of experience. This work explores machine learning models to address the key challenge of neighbor selection. Our goal is to predict the connection quality between two arbitrary viewers using features such as locality, access providers, operating systems, past CDN, and V2V throughput. The proposed solutions are validated using an A/B testing approach on our production system, demonstrating a significant improvement in key system metrics compared to the traditional locality-based methods. We observe 17% higher V2V throughput, 26% lower delay, 37% fewer lost chunks, 39% fewer re-buffering, and 20% fewer quality switches.