National Technical University of Athens (NTUA)
Μη-γραμμικά δυναμικά συστήματα και εύρωστη επεξεργασία και αναγνώριση φωνής
Abstract
dc:descriptionThe core research issue of this thesis lays in the area of characterization and analysis of speech signals. Inspiration for the initial steps of this research effort is bifold: from one hand there are indications on the existence of aerodynamic phenomena during speech production, like turbulence; at the same time there are recent advances in the theory of fractals on the quantification of the complexity of a set. Moreover the above two are inter-related. Additional motivation stems from, by considering the performance deterioration of speech recognition systems, especially when this is compared to the human ability. Given these directions, we propose speech processing methods inspired by the fractal geometry and dynamical systems theory. Analysis in this framework aims on the characterization of speech signals and on the subsequent application in speech recognition systems. The approach taken consists of the application of feature extraction methods that are related with the theory of fractal and generalized fractal dimensions. The basis of the approach is formed by reconstructing a phase space that is considered closer to the speech production system compared to the scalar speech signal. Further, by taking advantage of the reconstructed multidimensional geometry, the methods are extended for the case of noisy speech signals. The proposed methods are applied on classification experiments and on speech recognition experiments both in clean and noisy environmental conditions. The evaluation is conducted on well-known phonetic databases, and illustrates the prospects of the proposed methods for speech characterization. The results are promising especially for the cases of broad phoneme classification and noisy speech recognition. Moreover, in the framework of multiple feature fusion taking into account each feature’s uncertainty is shown to be fruitful for the problem of speech recognition.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2007
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Pitsikalis, Vasileios
- Πιτσικάλης, Βασίλειος
Subjects
dc:subject × 24- Φράκταλ (μορφοκλασματικές) διαστάσεις
- Μη γραμμικά δυναμικά συστήματα
- Ανακατασκευή χώρου φάσεων
- Ψυφιακή επεξεργασία σημάτων
- Ψηφιακή επεξεργασία φωνής
- Αυτόματη αναγνώριση φωνής
- Αυτόματη αναγνώριση προτύπων
- Αποθορυβοποίηση σημάτων φωνής
- Εύρωστη επεξεργασία και αναγνώριση φωνής
- Προσαρμοστική σύμμιξη πολλαπλών ροών πληροφορίας
- Fractal dimensions
- Non-linear dynamical systems
- Phase-space recostruction
- Digital signal processing
- Digital speech precessing
- Automatic pattern recognition
- Automatic speech recognition (ASR)
- Speech signals denoising
- Robust speech processing and recognition
- Adaptive fusion of multiple information streams
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Engineering and Technology
- Electrical Engineering, Electronic Engineering, Information Engineering
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/18282