Caracterización Automática del Llanto de Bebé para su Estudio con Modelos de Clasificación
As a part of a project that seeks to support early detection of pathologies in newborn babies, this thesis proposes a system of Automatic Infant Cry Recognition based on a characterization defined by the combination of acoustical features, which are obtained by different extraction techniques. Exper...
| Autor: | |
|---|---|
| Tipo de documento: | dissertação |
| Estado: | Versión aceptada para publicación |
| Data de publicação: | 2008 |
| País: | México |
| Recursos: | Instituto Nacional de Astrofísica, Óptica y Electrónica |
| Repositório: | Repositorio Institucional del INAOE |
| OAI Identifier: | oai:inaoe.repositorioinstitucional.mx:1009/393 |
| Acesso em linha: | http://inaoe.repositorioinstitucional.mx/jspui/handle/1009/393 |
| Access Level: | Acceso aberto |
| Palavra-chave: | info:eu-repo/classification/Clasificación/Classification info:eu-repo/classification/Reconocimiento de patrones/Pattern recognition info:eu-repo/classification/Extracción de características/Feature extraction info:eu-repo/classification/cti/7 info:eu-repo/classification/cti/33 info:eu-repo/classification/cti/3314 |
| Resumo: | As a part of a project that seeks to support early detection of pathologies in newborn babies, this thesis proposes a system of Automatic Infant Cry Recognition based on a characterization defined by the combination of acoustical features, which are obtained by different extraction techniques. Experiments were performed to recognize three types of cry: normal, pathological cry of hypo-acoustic (deaf) infants and asphyxia. The fact that the parameters have been derived from different spectral representation of the signal, suggests the possibility of raising different combinations of features to provide benefits to improve the representation of each type of crying, and consequently, increase the final recognition rate. In general, four characteristics extraction techniques were used: LPC (Linear Predictive Coding), MFCC (Mel Frequency Cepstral Coefficients), Intensity and Cochleograms. The original characteristic vectors were reduced through two methods like: LDA (Linear Discriminant Analysis), and a proposed method which is called, "Reduction by Statistics Operations". The combination of characteristics was carried out using the reduced characteristic vectors. The use of cochleograms to classify infant cry is one of the contributions of this thesis work. According to experiments, it was observed that cochleograms equalized, and in some cases improved the results obtained by techniques such as LPC or MFCC, which are widely used in speech recognition for their good results. Several tests were performed to validate the characterization. By applying traditional techniques such as ten-fold-cross-validation, results of an accuracy of 98.66% were achieved with vectors formed by the combination of four types of features. Other tests, which we call “individual tests” achieved results of 100% for the classification of the deaf class. Finally we defined a knowledge base for the classification of baby's cry considering the results and observations derived from this research. |
|---|