Publication: Diseño e implementación de un reconocedor de habla híbrido en Matlab
Loading...
Identifiers
Publication date
2013-06
Defense date
2013-07-09
Authors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
En este trabajo hemos abordado la implementación de las primeras etapas de
un reconocedor automático de habla en Matlab con una arquitectura híbrida
ANN/HMM (Artificial Neural Networks / Hidden Markov Models). En particular
se han implementado los módulos de adquisición, extracción de características
y parte del modelado acústico.
Para la adquisición, se han utilizado librerías estándar de matlab para poder
realizar la lectura de la base de datos ISOLET. Esta librería es ampliamente
conocida en el área del reconocimiento automático de habla.
Posteriormente, se ha utilizado la librería Voicebox para obtener los
coeficientes MFCC (Mel Frequency Cepstral Coefficients) así como los
coeficientes dinámicos correspondientes. Además, se ha añadido un
procedimiento para construir un contexto para cada uno de los vectores de
parámetros.
Por último, se ha realizado una búsqueda y posterior selección de una librería
matlab para la implementación de MLP (Multi-Layer Perceptrons) con los
requisitos necesarios para su posterior integración con los HMM. Así
finalmente, se ha implementado un módulo de estimación de las probabilidades
a posteriori de los vectores anteriormente descritos dada las 28 posibles clases
de fonemas de nuestro entorno de experimentación.
In this project, we have implemented the first steps of an Automatic Speech Recognizer (ASR) in Matlab employing a hybrid ANN/HMM (Artificial Neural Networks / Hidden Markov Models) scheme. In particular, modules for the acquisition, feature extraction and acoustic modeling (partly) have been implemented. Standard Matlab libraries have been employed for the acquisition module enabling the sequential reading of the well known ISOLET database. Then, Voicebox, a library specifically designed for speech processing, has been employed for the computation of MFCC (Mel Frequency Cepstral Coefficients). Besides, a procedure for the construction of an acoustical context for each of the feature vectors has been included. Finally, a search process and subsequent selection of a neural network matlab library for the implementation of MLP (Multi-Layer Perceptrons) with the requirements needed for the foreseen integration into HMM has been carried out. Lastly, the 'a posteriori' probabilities estimation module for each of the feature vectors previously described given the 28 possible phonetic labels of our experimental testbed was implemented.
In this project, we have implemented the first steps of an Automatic Speech Recognizer (ASR) in Matlab employing a hybrid ANN/HMM (Artificial Neural Networks / Hidden Markov Models) scheme. In particular, modules for the acquisition, feature extraction and acoustic modeling (partly) have been implemented. Standard Matlab libraries have been employed for the acquisition module enabling the sequential reading of the well known ISOLET database. Then, Voicebox, a library specifically designed for speech processing, has been employed for the computation of MFCC (Mel Frequency Cepstral Coefficients). Besides, a procedure for the construction of an acoustical context for each of the feature vectors has been included. Finally, a search process and subsequent selection of a neural network matlab library for the implementation of MLP (Multi-Layer Perceptrons) with the requirements needed for the foreseen integration into HMM has been carried out. Lastly, the 'a posteriori' probabilities estimation module for each of the feature vectors previously described given the 28 possible phonetic labels of our experimental testbed was implemented.
Description
Keywords
Reconocimiento de voz, MATLAB (Programa de aplicación), Redes neuronales