Catálogo de publicaciones - tesis
Título de Acceso Abierto
Técnicas de aprendizaje maquinal para separación ciega de fuentes sonoras con aplicación al reconocimiento automático del habla
Leandro Ezequiel Di Persia Diego Humberto Milone Juan Cosseau Carlos Muravchik Juan Carlos Gómez Leonardo Luis Giovanini Masuzo Yanagida
acceptedVersion.
Resumen/Descripción – provisto por el repositorio digital
In the last decades a new problem related to machine learning and signal processing has emerged in many disciplines: the blind source separation problem. The blind source separation technique aims to segregate the sources that contribute to some variation of a physical quantity, given a set of measurements of the global variation produced by all sources at a time. One particular application of the blind source separation methods is the Automatic Speech Recognition, which can be defined as the task of determining the text that corresponds to a given spoken utterance. This kind of systems have reached a maturity point but they still suffer from a strong drawback: they cannot adequatelly manage the existence of noise or competing sources in the input. This doctoral dissertation presents several advances in the technique of audio source separation in reverberat conditions, using independent component analysis in the time-frequency domain. Three methods were developed in order to produce a better quality of separation and, at the same time, to reduce the processing times. The proposed algorithms were evaluated under realistic conditions such as different environments and different kind and power of competing sources. For this purpose we used two evaluation alternatives, objective quality measures of the resulting signal and the performance in the application of interest, that is, automatic speech recognition. The results for the different approaches show the possibility of getting through the dilemma between resulting quality and requiered processing time, converging to a very fast and high quality separation method.Palabras clave – provistas por el repositorio digital
Blind source separation; Independent component analysis; Reverberation; Ambient noise; Robust speech recognition; Objective quality evaluation; Separación ciega de fuentes sonoras; Análisis de componentes independientes; Reverberación; Ruido del ambiente; Reconocimiento robusto del habla; Evaluación objetiva de calidad
Disponibilidad
| Institución detectada | Año de publicación | Navegá | Descargá | Solicitá |
|---|---|---|---|---|
| No requiere | 2009 | Biblioteca Virtual de la Universidad Nacional del Litoral (SNRD) |
|
Información
Tipo de recurso:
tesis
Idiomas de la publicación
- inglés
País de edición
Argentina
Fecha de publicación
2009-03-26
