Catálogo de publicaciones - tesis
Título de Acceso Abierto
Modelos de factorización en matrices no negativas para procesamiento de audio
Francisco Javier Ibarrola Leandro Ezequiel Di Persia Hugo Aimar Ana Georgina Flesia Gastón Schlotthauer Rubén Daniel Spies
acceptedVersion.
Resumen/Descripción – provisto por el repositorio digital
Along the technological advances regarding portable electronic devices, over the last years the need for improving human machine interactions through speech has arise. One of the main challenges within this contexts has to do with the fact that a recording device does not register the target signal directly, but distorted with noise, echoes and other sound sources. In order to model these phenomena in a computationally efficient way and to gain interpretability, we can make use of nonnegative matrix factorization (NMF) techniques, that allow for a piecewise and purely additive representation of the data. Nevertheless, the classical NMF approaches present drawbacks associated to the lack of uniqueness in the representations. Furthermore, the current optimization methods are based on iterative, multiplicative algorithms that are not immediately adaptable to models contemplating certain time or frequency relations between their elements. In this thesis, new NMF methods are developed in order to tackle the problems of dereveberation and speech source separation. Through a Bayesian approach, and the associated penalization functions, certain characteristics are imposed over the elements constituting the model, which allows sorting out some classical difficulties within this context, associated to the lack of uniqueness and correlation in representations. Also, optimization algorithms are developed in order to tackle the minimization problems associated to the built functional, that allow for incorporating certain types of penalizers that due to their characteristics cannot be defined element by element.Palabras clave – provistas por el repositorio digital
Dereverberation; Source separation; Machine learning; Signal processing; Dereverberación; Separación de fuentes sonoras; Aprendizaje maquinal; Procesamiento de señales
Disponibilidad
| Institución detectada | Año de publicación | Navegá | Descargá | Solicitá |
|---|---|---|---|---|
| No requiere | 2019 | Biblioteca Virtual de la Universidad Nacional del Litoral (SNRD) |
|
Información
Tipo de recurso:
tesis
Idiomas de la publicación
- español castellano
País de edición
Argentina
Fecha de publicación
2019-10-08
