Back to search

National Technical University of Athens (NTUA)

Σύνθεση φωνής με υπολογιστική αεροδυναμική ανάλυση του ανθρωπινού ηχητικού σωλήνα και σύγκριση με κλασσικές μεθόδους

Abstract

dc:description

The thesis studies the non-linear phenomena in the vocal tract and glottal source during speech production, so as to improve speech applications, such as speech synthesis and speech recognition. The thesis’ contribution is threefold. Firstly, we formally analyze non-linear phenomena as reflected in the speech signal, secondly, we propose novel acoustic features for speech recognition, and thirdly, we describe a complete state-of-the-art text-to-speech system .The analysis of non-linear phenomena is indirectly performed in the speech signal with the aid of the non-linear AM-FM speech model, which captures the non-linear phenomena as amplitude and frequency modulations. We define Amplitude Modulation Index and Frequency Modulation Index that measure the amplitude and frequency modulation percentages respectively. We estimate the modulation indexes on a large speech corpus, and analyze the correlations with various factors. Fundamental frequency is an important factor for the appearance of non-linear phenomena, while gender and manner of articulation are strong correlates. We then exploit the analysis’ results to develop a better acoustic front-end for speech recognition. Specifically, we propose a novel acoustic representation based primarily on frequency features that has significant advantages compared to standard representations. The proposed frequency acoustic features can be estimated in either time or frequency domain utilizing corresponding filter banks. The time domain estimation is computed as the averaged instantaneous frequency normalized by the energy, while the frequency domain estimation is equal to the first spectral moment also normalized by the energy. Moreover, we find the optimal algorithmic parametrization for the computation of the proposed features regarding the filter bank setup, and the decorrelation of the feature vector. The evaluation shows that the proposed feature has superior performance than alternative standard front-ends, both for clean and noisy recording conditions. Finally, we present a complete text-to-speech system that incorporates state-of-the-art techniques. Specifically, the system is based on the selection of elementary speech units from a large speech corpus, namely it is a unit selection system. Moreover, it incorporates a sophisticated text preprocessing front-end tailored to the rich morphology of the Greek language. Furthermore, we propose a novel technique for the optimization both in terms of computational load and storage resources.

Degree

thesis:*
Grantor dc:publisher
National Technical University of Athens (NTUA)
Year dc:date
2010

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Τσιάκουλης, Πύρρος
  • Tsiakoulis, Pirros

Subjects

dc:subject × 16

Rights

Language dc:language
gre

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/24302

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Τσιάκουλης, Πύρρος; Tsiakoulis, Pirros. Σύνθεση φωνής με υπολογιστική αεροδυναμική ανάλυση του ανθρωπινού ηχητικού σωλήνα και σύγκριση με κλασσικές μεθόδους. National Technical University of Athens (NTUA), 2010. http://hdl.handle.net/10442/hedi/24302