Algorithmes évolutionnaires pour l'étude de la robustesse des systèmes de reconnaissance de la parole

Abstract : Automatic speech recognition systems are becoming ever more common and are increasingly deployed in more variable acoustic conditions, by very different speakers. So these systems, generally conceived in a laboratory, must be robust in order to provide optimal performance in real situations. This ph-D explores the possibility of gaining robustness by designing speech recognition systems able to auto-modify in real time, in order to adapt to the changes of acoustic environment. As a starting point, the adaptive capacities of living organisms were considered in relation to their environment. Analogues of these mechanisms were then applied to automatic speech recognition systems. It appeared to be interesting to imagine a system adapting to the changing acoustic conditions in order to remain effective regardless of its conditions of use. Initially, the speech recognition system itself was adapted to various environments. Its capacity to adapt to the changes of acoustic conditions was studied, using a local approach (by retro-propagation of the gradient) and a global solution (by evolutionary algorithms), in order to find an optimal system. Secondly, the specific aspects of the system's input data processing were examined. A projection base adapted to each environment was sought, based on a principal component analysis of acoustic data, using evolutionary algorithms to set the system's knowledge of acoustic conditions. A simulation platform was set up, to allow the evolution of populations of recognition systems. Results obtained show that on average the hybridization of the evolutionary algorithms and traditional techniques of recognition improves the performance of the speech recognition system.
Document type :
Theses
Complete list of metadatas

Cited literature [36 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-00004850
Contributor : Thèses Imag <>
Submitted on : Wednesday, February 18, 2004 - 4:41:44 PM
Last modification on : Thursday, April 11, 2019 - 12:04:14 PM
Long-term archiving on : Tuesday, September 7, 2010 - 5:00:35 PM

Identifiers

  • HAL Id : tel-00004850, version 1

Collections

UJF | IMAG | UGA

Citation

Anne Spalanzani. Algorithmes évolutionnaires pour l'étude de la robustesse des systèmes de reconnaissance de la parole. Autre [cs.OH]. Université Joseph-Fourier - Grenoble I, 1999. Français. ⟨tel-00004850⟩

Share

Metrics

Record views

345

Files downloads

761