Thèse soutenue

FR
Auteur / Autrice : Kenko Ota
Direction : Philippe VanheegheEmmanuel DuflosMasuzo Yanagida
Type : Thèse de doctorat
Discipline(s) : Informatique industrielle et automatique
Date : Soutenance en 2008
Etablissement(s) : Ecole Centrale de Lille en cotutelle avec Kyōto Doshisha daigaku

Résumé

FR  |  
EN

Speech recognition technology reaches almost a practical level if we use a close contact microphone in quiet environments. However, in case microphones are located at a distant position from a speaker, it is necessary to develop noise reduction and dereverberation techniques. A technique for reducing obstructive sounds emitted by the target apparatus to be controlled is proposed. The proposed system uses harmonic structure of voiced segments which conventional ANCs does not directly take into account. A new dereverberation technique considering the frequency characteristics on reflective surfaces is also proposed. Over-subtraction occurs in conventional dereverberation in case of flat frequency characteristics. So, it is required to estimate the actual reverberation time assuming the frequency characteristics of reflection. Proposed is a single channel blind dereverberation technique using auto-correlation functions on the time sequences of frequency components. A technique to escape from the permutation problem which appears in frequency-domain Independent Component Analysis (ICA) is also proposed : the Multi-bin ICA (MB-ICA). Finally, a technique to estimate speech spectrum using a particle filter with a single microphone is proposed. This technique consists in estimating noise and speech spectra using a model based on Dirichlet Process Mixture (DPM) instead of the Gaussian Mixture Model (GMM). It is thus expected to develop a method to estimate the spectrum adaptively