SPEECH RECOGNITION AND SYNTHESIS

Author: MONTORO MANRIQUE GERMÁN.
Year: 2004.
University: AUTÓNOMA DE MADRID.
Place of defense: ESCULA POLITÉCNICA SUPERIOR.
Place of preparation: ESCUELA POLITÉCNICA SUPERIOR.

Summary: This paper begins by showing the main features that present intelligent environments and interfaces oral dialogues that can be used in them. He then introduced a proposal for ambient intelligence followed by solutions for the development of oral interfaces in these spaces. The work detailing the functionality of the system of dialogues implemented, in accordance with the requirements laid down for this type of environments. The approach taken pays special importance to the need for interfaces can be created automatically and adapt to the dynamic characteristics of intelligent environments. Finally, we show the results of the assessment with user interface proposal in a real environment.

Author: GÓRRIZ SÁEZ JUAN MANUEL.
Year: 2005.
University: GRANADA.
Place of defense: ESCUELA TÉCNICA SUPERIOR DE INGENIERÍAS INFORMÁTICA Y DE TELECOMUNICACIÓN.
Place of preparation: ESCUELA TÉCNICA SUPERIOR DE INGENIERÍAS INFORMÁTICA Y DE TELECOMUNICACIÓN.

Summary: This paper presents new developments in the field of voice activity detection (VAD, the English "Voice Activity Detection") for its application to Speech Recognition robust in noisy environments. The new voice activity detector (VADs) are based on different methodologies: 1-Tests based on statistical averages biespectrales on the grid in the domain bi-frecuencia. 2-Tests based on the statistical likelihood ratio (LRT) magnitudes biespectrales integrated. 3-Cluster Analysis Modeling space noise and development of a rule-based decision divergence cluster. 4-support vector machines (SVMs) applied to the classes of SNRs in subbands energy. The performance of VADs proposed is higher success rate for detection of a false alarm because, when compared with VADs standard, such as those of ITU-T G.729, AMR and ETSI ETSI GSM AFE, and the recently published using databases most representative ETSI like AURORA2 & 3, and as part of a system of automatic recognition (ASR), significantly improves the rate of appreciation of speech in a noisy environment.

Author: Fiérrez Aguilar Julián.
Year: 2005.
University: POLITÉCNICA DE MADRID.
Place of defense: E.T.S. DE INGENIEROS DE TELECOMUNICACION.
Place of preparation: ETSI Telecomunicación.

Summary: It's called biometric process that allows associating with an individual identity automatically, using any personal characteristic inherent to it. The goal in the biometric verification is deciding between two classes, client or imposter. To this end there are a number of physiological patterns that can be considered as: fingerprint, iris, face, the geometry of the fingers and / or hand, or the palmar fingerprint, voice, writing and handwritten signature , or gait (ongoing), ... This thesis focuses on the combination of multiple biometric features to overcome some of the limitations of individual traits, in what is known as a multi-modal biometric system. Most strategies for multimodal fusion in the literature are based on the merger of scores provided by the individual systems. In most cases it is assumed that these schemes combination remain stationary during the verification phase. This thesis focuses on the study of schedules tailored merger during the verification phase and raises new architectures merger biometric authentication adapted to the different users registered in the system, such as the quality of biometric input signals; resulting in merger schemes tailored for multimodal biometric authentication. These schemes are also other application scenarios combining different sources of information and not necessarily different biometric features, such as the merger multi-sensor, multi-algoritmo, multi-unidad, and repeated units. The term 'adjusted' in this thesis deals with schemes trained using general information, such as a reference set of users, and adjusted well to consider specific information of the user in question either the quality of the sample biometric entry. In this sense, user-dependent approaches presented in this thesis are originals. Similar approaches in the literature or are trained only with the generic information, or are trained to information including user, but do not use both. They are also original adaptation of the work taking into account the quality biometrics. The techniques of fusion in the majority of cases are presented on both implementations based on statistical models as in Support Vector Machine (SVM). Experimental studies of the thesis is based on common practices for evaluating performance in multimodal authentication systems, and describes and presents the multimodal biometric databases used. The experimental results beginning with the implementation of the fusion techniques adapted to three problems biometric authentication unimodal multi-algoritmo based on written signature, voice and fingerprint, respectively. Realizándose a comparative study of the schemes tailored presented. Finally explores merger schemes tailored to both user and the quality of the input signals at its two proposed versions, which are based on statistical models or SVM. It demonstrates the superiority of the schemes tailored to the user compared to traditional methods independent user, or dependent user without use of general information. In the comparative study demonstrates the superiority of the scheme based on SVM when the number of samples is reduced training, and the statistical approach when the number of samples is high. In the event of a merger is dependent on quality improvement shows provided to take into account the quality in both approaches, especially in the scheme based on SVM.