A text-to-speech system is a complex system consisting of several different modules such as grapheme-to-phoneme conversion, articulatory and prosodic modelling, voice modelling etc. This dissertation is aimed at the creation of the initial part of a text-to-speech system, i.e. the grapheme-to-phoneme conversion, designed for Swedish. The problem area at hand is the conversion of orthographic text into a phonetic representation that can be used as a basis for a future complete text-to speech system. The central issue of the dissertation is the grapheme-to-phoneme conversion and the elaboration of rules and algorithms required to achieve this task. The dissertation aims to prove that it is possible to make such a conversion by a rule-based algorithm with reasonable performance. Another goal is to find a way to represent phonotactic rules in a form suitable for parsing. It also aims to find and analyze problematic structures in written text compared to phonetic realization. This work proposes a knowledge-based grapheme-to-phoneme conversion system for Swedish. The system suggested here is implemented, tested, evaluated and compared to other existing systems. The results achieved are promising, and show that the system is fast, with a high degree of accuracy.

The purpose of this study is to investigate sixth form students’ pronunciation, and their exposure to English during their English lessons in school. The focus of the study is to investigate whether or not the students have problems with pronouncing the contrastive consonant phonemes that do not exist, or are rarely used in the Swedish language (i.e /z/). In order to investigate the students’ pronunciation, questionnaires were handed out, followed by a reading exercise that was recorded. Also, a questionnaire was handed out to the students’ teachers in order to investigate their thoughts about the importance of teaching pronunciation. The participating students and teachers in this essay were chosen from a school in the south-west part of Sweden. The results in this essay show that the majority of the students participating had difficulties pronouncing the English consonant phonemes which do not exist, or are rarely used, in Swedish i.e /z/, /tʃ/ and /dʒ/. Furthermore, the results in this essay show that the students are more likely to pronounce English words with consonant phonemes similar to those used in Swedish.

Please read the abstract on page 112 in the dissertation / Dissertation (MA)--University of Pretoria, 2009. / African Languages / unrestricted

During spoken language comprehension, listeners transform continuous acoustic cues into categories (e.g. /b/ and /p/). While longstanding research suggests that phoneme categories are activated in a gradient way, there are also clear individual differences, with more gradient categorization being linked to various communication impairment like dyslexia and specific language impairments (Joanisse, Manis, Keating, & Seidenberg, 2000; López-Zamora, Luque, Álvarez, & Cobos, 2012; Serniclaes, Van Heghe, Mousty, Carré, & Sprenger-Charolles, 2004; Werker & Tees, 1987). Crucially, most studies have used two-alternative forced choice (2AFC) tasks to measure the sharpness of between-category boundaries. Here we propose an alternative paradigm that allows us to measure categorization gradiency in a more direct way. We then use this measure in an individual differences paradigm to: (a) examine the nature of categorization gradiency, (b) explore its links to different aspects of speech perception and other cognitive processes, (c) test different hypotheses about its sources, (d) evaluate its (positive/negative) role in spoken language comprehension, and (e) assess whether it can be modified via training. Our results provide validation for this new method of assessing phoneme categorization gradiency and offer valuable insights into the mechanisms that underlie speech perception.

Speech processing systems have been developed for many years but the integration into devices had started with the deployment of the modern powerful computational systems. This dissertation thesis deals with development of the keyword detection system in speech data. The proposed detection system is based on the Large Margin and Kernel methods and the key part of the system is phoneme classifier. Two hierarchical frame-based classifiers have been proposed -- linear and non-linear. An efficient training algorithm for each of the proposed classifier have been introduced. Simultaneously, classifier based on the Gaussian Mixture Models with the implementation of the hierarchical structure have been proposed. An important part of the detection system is feature extraction and therefor all algorithms were evaluated on the current most common feature techniques. A part of the thesis technical solution was implementation of the keyword detection system in MATLAB and design of the hierarchical phoneme structure for Czech language. All of the proposed algorithms were evaluated for Czech and English language over the DBRS and TIMIT speech corpus.

To address the needs of the growing number of Spanish-English bilingual children in the United States, Nonword Repetition (NWR) tasks were created to reduce testing bias in the assessment and diagnosis of children with developmental language disorder (DLD). Several studies have shown promising results in the use of NWR tasks; however, fewer studies have addressed questions such as the use of different scoring methods or analyzing error patterns. Thus, this study was conducted to address these gaps in the research. An English and a Spanish NWR task were administered to 26 Spanish-English bilingual school aged children (6;0-9;4). Two different scoring methods (percent phoneme correct and whole word scoring) were compared for diagnostic accuracy and the types and frequency of errors were analyzed. Both scoring methods showed statistically significant differences between groups (participants with DLD and those with typically developing language). Whole word scoring in Spanish had the best diagnostic accuracy, according to sensitivity, specificity, and likelihood ratio measures. However, due to the small number of nonwords that any participant repeated correctly, this may not be a clinically practical scoring method. The Spanish NWR task was a better measure than the English NWR task in identifying children with DLD, suggesting that Spanish NWR could be used to assess DLD in bilingual children. Participants with DLD produced more consonant, vowel, substitution, and omission errors than those with typically developing language. There was no difference between groups for addition errors. Significantly more omission errors were made in Spanish, likely due to the longer nonwords. The longer nonwords may be key in distinguishing between typically developing children and those with DLD. These results have the potential to inform future clinical practices in selecting, scoring, and analyzing NWR tasks.

Phonological processing, the ability to recognize and manipulate the sounds of one's native language, is an essential linguistic skill. Deficits in this skill may lead to decreased social, educational, and financial success (Kraus & White-Schwoch, 2019). Additionally, phonological disorders have been shown to be highly variable and individualized (Bellon-Harn & Cradeur-Pampolina, 2016) and therefore difficult to treat effectively. A better understanding of the neural underpinnings of phonological processing, including the underlying skill of phonemic discrimination, could lead to the development of more individualized and effective intervention. Several studies, some using quantitative electroencephalography (qEEG) and others using functional magnetic resonance imaging (fMRI), have been conducted to investigate these neural underpinnings. When considering the relative strengths and weaknesses of qEEG and fMRI, the scientific community has traditionally believed qEEG to be excellent at determining when brain activity occurs (temporal resolution), but to have limited abilities in determining where it occurs (spatial resolution). On the other hand, the reverse is believed to be true for fMRI. However, the spatial resolution of qEEG has improved over recent decades and some studies have reached levels of specificity comparable to fMRI. This thesis provides an abbreviated meta-analysis determining the accuracy and consistency of source references, or areas where brain activation is determined to originate from, in qEEG studies evaluating phonemic discrimination. Nineteen experiments were analyzed using the Comprehensive Meta-Analysis software. A study's event rate was defined as the number of times an anatomical area was coded as a source reference, divided by the participants in the study. Results show that each of these experiments had relatively low event rates, culminating into a summary event rate of 0.240. This indicates that qEEG does not provide source references that are as accurate or consistent as fMRI. This meta-analysis concludes that although there is research suggesting qEEG may have developed to be comparable to fMRI in spatial resolution, this is not supported in the analysis of qEEG studies focused on phonemic discrimination.

In Czechia, teaching L2 English pronunciation is underestimated, which might lead to potential production slips and affect the intelligibility of the learners. This study investigates accuracy in L2 English pronunciation of checked vowels (/ʌ, ɛ, ɪ, ɒ, ʊ, æ/) as produced by Czech students. I applied cross-sectional approach and analysed participants from two distinct school classes who are divided by an 8-year study period. The pronunciation accuracy is analysed with the help of the computer program Praat, which generates frequencies for both Czech and English vowel phonemes produced by the participants reading out loud a set of monosyllabic citation words. The formant frequencies are later compared to a reference set of SSBE frequency values, which represents the targeted native language model. The results of this work show that the L2 English values are influenced by the students’ inventories of L1 Czech, leading to inaccuracy in production of the phonemes /ʌ, ɛ, ɪ, ɒ, ʊ/ and setting constraints to the L2 /æ/ acquisition.

This dissertation suggests a new approach to reading instruction and phonetic control in materials development. The central feature of this approach is its reliance on phoneme-grapheme correspondence data. Specifically, the products of this dissertation were (1) the development of relative frequency tables depicting the various phonemes produced by a given grapheme structure, (2) the development of five different levels of phonetic sophistication, (3) the classification of 974 words as regular or irregular in accordance with each level of sophistication, and (4) the phonetic rewriting of three existing short stories to make them phonetically regular. One of the phonetically rewritten stories was given to an existing group of sixth grade students along with its original form. The students performed with superior reading speed and equal comprehension on the phonetically rewritten story. This difference was significant at the .01 level.

This diploma thesis discusses modification of a speech rate. The PSOLA (Pitch Synchronous OverLap Add) method was used for the rate modification. This algorithm works in time domain. Another method -- phase vocoder, which works in frequency domain is also presented in an overview. This thesis extends the PSOLA method with a phoneme recognition, which allows for better understandability of the speech output by considering characteristics of the phonemes beeing pronounced. To examine this proposed method, an application connecting PSOLA and a phoneme recognizer was developed.