State-of-the-art automatic speech recognition (ASR) systems are based on probabilistic modelling of the speech signal using Hidden Markov Models (HMM). These models lead to the best recognition performances in ideal "lab" conditions or for easy tasks. However, in real word conditions of speech processing, the performances of HMM-based ASR systems can decrease drastically and their use becomes very limited. For this reason, the conception of robust and viable ASR systems has been a tremendous scientific and technological challenge in the field of ASR for the last decades. The scope of this thesis is to address this challenge by attacking the core of the problem what we believe is the robustness in speech modelling. Precisely, our strategy is to conceive ASR systems for which robustness relies on: * the fidelity and the flexibility in speech modelling rather than (ad-hoc) tuning of HMMs, * a better exploitation of the information contained in the available statistical data.
展开▼
