Mircea Giurgiu

NEURAL NET WORD REPRESENTATIONS FOR PHRASE-BREAK PREDICTION WITHOUT A PART OF SPEECH TAGGER

The use of shared projection neural nets of the sort used in language modelling is proposed as a way of sharing parameters between multiple text-to-speech system components. We experiment with pretraining the weights of such a shared projection on an auxiliary language modelling task and then apply the resulting word representations to the task of phrase-break prediction. Doing so allows us to build phrase-break predictors that rival conventional systems without any reliance on conventional knowledge-based resources such as part of speech taggers.

MediaWiki interoperability framework for multimedia digital resources

The success of the collaborative web-based MediaWiki platform, widely used in several projects to exchange knowledge created a new idea to use this system as a low-tech interoperability and repository layer for data providers, end users, developers and project partners. Facilitating the acquisition of knowledge for multimedia digital resources is a task that usually requires special purpose interfaces with which users are not familiar. The method effectively enables data providers to publish their metadata about multimedia content in the field of biodiversity in a push-operation to a metadata repository through a familiar interface like MediaWiki templates. The workflow then involves a procedure for automatic metadata harvesting into Fedora Commons repository, combined with the automatic creation of repository reports written to wiki pages in order to ensure a feedback to the data providers and end users. Models, techniques, standards and protocols used in the KeyToNature project make MediaWiki a layered candidate in achieving interoperability at the syntactic and semantic level with a low technological entry barrier.

Romanian language statistics and resources for text-to-speech systems

This paper introduces a series of results and experiments used in the development of a Romanian text-to-speech system, focusing on text statistics. We investigate the presence of several linguistic units used in text-to-speech systems, from phonemes to words. The text corpus we used, News-Romanian (News-RO) comprises 4500 newspaper articles. A subset of it, around 2500 sentences represents the Romanian Speech Synthesis (RSS) recorded speech database. The results offer an important insight to how should a speech database be designed. We also describe the methods used in the development of a 50,000 words Romanian lexicon with phonetic transcription and accent positioning. Such a lexicon is useful in machine learning algorithms of the front-end part of a text-to-speech system. As an addition we study the use of Maximal Onset Principle for Romanian syllabification.

Automatic transcription and speech recognition of Romanian corpus RO-GRID

The results reported in this paper assess the ability of Hidden Markov Model (HMM) based method to generate accurate and reliable automatic phone-level transcriptions for a small vocabulary speech corpus such as RO-GRID. The system requires only orthographic transcription of the target corpus, and can be bootstrapped from models trained just on few amount of data in the transcribed corpus. For this purpose, an automatic time-aligned phone transcription toolbox has been developed and tested on the Romanian corpus and also validated on an English corpus. The quality of transcriptions is judged by evaluating the statistical parameters of the error between the automatic and manual transcription. The transcriptions generated from the most reliable system deviate from the average manual transcription by an average of 20 ms. The system is also able to convert the generated transcription from HTK format into PRAAT format for further manipulation of the speech signal.

A superpositional model applied to F0 parameterization using DCT for text-to-speech synthesis

This paper addresses the idea of the superpositional model based on the DCT (Discrete Cosine Transform) parameterization of the F0 contours. We examine the capacity of the DCT coefficients to estimate the fast variations in the F0 contour at syllable level and also the overall trend of the phrase level. The method determines the coefficients at syllable level, based on the subtraction of the estimated phrase level contour from the original one; thus considering that the syllable has an additive prosodic effect over the phrase level. We also compare the use of 3 different decision and regression tree algorithms for DCT coefficients clustering and prediction. Additional features are selected based on a greedy stepwise without backtracking feature selection method. The results support the proposed method through low average square errors and little or no perceivable errors in the synthesized speech.

Implementation of a security layer for the SSL/TLS protocol

This article provides a systematic approach on how network applications with support for secure SSL/TLS protocols work. In order to illustrate the functionality and the development of such an application the framework functions and procedures will be exemplified. The sample application proposed is targeted for the Windows OSs and implements the open-source library OpenSSL using managed .NET code.

Voice-related quality of life results in laryngectomies with today's speech options and expectations from the next generation of vocal assistive technologies

Objectives: To assess the voice handicap, the satisfaction with todays voice assisting methods and to identify the needs that should be addressed by new vocal assistive technologies for aphonic patients. Materials and Methods: We conducted a prospective study on two samples of patients with total laryngectomy and submitted to speech therapy. Voice Handicap Index (VHI) questionnaires and qualitative (focus-groups) and quantitative (online surveys) methods were used. Results: Analysis of the VHI total revealed that the esophageal and electrolarynx speakers had a moderate voice handicap, while tracheoesophageal speakers and patients without vocal rehabilitation had a severe handicap. Interview and survey data indicated that these patients have many needs which are unmet by available rehabilitation methods. Conclusions: These results point out the necessity to improve current vocal assistive methods and to develop better technologies that could increase the quality of life of this patients.

Phonetic segmentation of speech using STEP and t-SNE

This paper introduces a first attempt to perform phoneme-level segmentation of speech based on a perceptual representation - the Spectro Temporal Excitation Pattern (STEP) - and a dimensionality reduction technique - the t-Distributed Stochastic Neighbour Embedding (t-SNE). The method searches for the true phonetic boundaries in the vicinity of those produced by an HMM-based segmentation. It looks for perceptually-salient spectral changes which occur at these phonetic transitions, and exploits t-SNEs ability to capture both local and global structure of the data. The method is intended to be used in any language and it is therefore not tailored to any particular dataset or language. Results show that this simple approach improves segmentation accuracy of unvoiced phonemes by 4% within a 5 ms margin, and 5% at a 10 ms margin. For the voiced phonemes, however, accuracy drops slightly.

Evaluation of sentiment polarity prediction using a dimensional and a categorical approach

In this paper we evaluate two approaches for predicting the sentiment polarity of an utterance. The first method is based on a 3-dimensional model which takes into account text expressiveness in terms of valence, arousal and dominance. The second one determines the words semantic orientation according to Chi-square and Relevance factor statistic metrics. We describe the general flow of the methods and their extracted features, as well as their predictability potential using different machine learning algorithms, Naïve Bayes, SVM and C4.5. The evaluation is performed on four emotional datasets: Semeval 2007 “Affective Text”, ISEAR (International Survey on Emotional Antecedents and Reactions), childrens fairy-tales and a movie review dataset. The results show a high correlation of the prediction performance with the database content, as well as to the average number of words within the classified text instances.

Semantic MediaWiki interoperability framework from a semantic social software perspective

This paper presents two collaborative Social-Software-driven approaches for the interoperability of multimedia resources used in KeyToNature project. The first approach, using MediaWiki as a low level interoperability framework is presented in our previous works. The second one, Semantic MediaWiki interoperability framework for multimedia resources is presented in this paper, and is still in progress. We are arguing that different approaches are needed, depending on the context and intention of multimedia resource use.