Publication: Reconocimiento automático de habla con adaptación al género y al locutor
Loading...
Identifiers
Publication date
2010-12
Defense date
Authors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
El presente proyecto fin de carrera tiene como objetivo mejorar las prestaciones de un reconocedor automático de habla continua en castellano, adaptando sus características al género y al locutor. ,- Para ello, como sistema base se ha desarrollado un reconocedor automático de habla continua para la base de datos MICROAES, utilizando el conjunto de herramientas de HTK con el que se realiza un modelado basado en HMMs. Para observar el efecto que provoca introducir información sobre el género en el reconocimiento, se ha utilizado una adaptación MAP al género sobre los modelos acústicos previamente entrenados. Además, se ha implementado un clasificador de género con el fin de realizar automáticamente la elección del género correspondiente a cada locución. Finalmente, la adaptación MAP se ha comparado con el entrenamiento por separado de los modelos dependientes del género. Para analizar la influencia producida por la incorporación de la información del locutor, se ha planteado el uso de dos técnicas de adaptación diferentes, MAP y MLLR. Por último, se ha evaluado el uso conjunto de la adaptación al género y al locutor, utilizando una combinación de las técnicas de adaptación mencionadas anteriormente. --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
This Final Degree Project aims to improve the recognition rate of an automatic speech recognition system in Spanish, by adapting its characteristics to gender and speaker variability. For that, as a baseline system; an automatic cQntinuous speech recognizer has been developed for the MICROAES database, using the HTK toolkit to perform a modeling based on HMMs. To observe the effect caused by introducing gender information in the recognition system, a gender MAP adaptation of the acoustic models previously trained has been used. Furthermore, a gender classifier to automatically perform the choice of gender for each utterance has been implemented. Finally, the MAP adaptation technique has been compared to the separate training of gender dependent models. To analyze the influence of the incorporation of the speaker information, two different adaptation techniques, MAPand MLLR, have been considered. Finally, the joint use of gender and speaker adaptation has been evaluated using a combination of the adaptation techniques previously mentioned.
This Final Degree Project aims to improve the recognition rate of an automatic speech recognition system in Spanish, by adapting its characteristics to gender and speaker variability. For that, as a baseline system; an automatic cQntinuous speech recognizer has been developed for the MICROAES database, using the HTK toolkit to perform a modeling based on HMMs. To observe the effect caused by introducing gender information in the recognition system, a gender MAP adaptation of the acoustic models previously trained has been used. Furthermore, a gender classifier to automatically perform the choice of gender for each utterance has been implemented. Finally, the MAP adaptation technique has been compared to the separate training of gender dependent models. To analyze the influence of the incorporation of the speaker information, two different adaptation techniques, MAPand MLLR, have been considered. Finally, the joint use of gender and speaker adaptation has been evaluated using a combination of the adaptation techniques previously mentioned.
Description
Keywords
Reconocimiento automático del habla, HMMs, Adaptación, MAP, MLLR, Clasificador de género, Automatic speech recognition, Adaptation, Gender classifier