Age prediction by voice using deep learning

One of the main topics in artificial intelligence is the speech characterization. Moreover, it is a field of study with the minimal scope when the Catalan language is involved in. In this project, we try to perform an age classification by decades firstly in the Catalan CommonVoice Dataset and then...

Descripción completa

Detalles Bibliográficos
Autor: Linde Martínez, David
Tipo de recurso: tesis de maestría
Fecha de publicación:2023
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/386585
Acceso en línea:https://hdl.handle.net/2117/386585
Access Level:acceso abierto
Palabra clave:Deep learning
Artificial intelligence
Automatic speech recognition
deep learning
artificial intelligence
voice
age
Aprenentatge profund
Intel·ligència artificial
Reconeixement automàtic de la parla
Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial
Àrees temàtiques de la UPC::Enginyeria de la telecomunicació::Processament del senyal::Processament de la parla i del senyal acústic
Descripción
Sumario:One of the main topics in artificial intelligence is the speech characterization. Moreover, it is a field of study with the minimal scope when the Catalan language is involved in. In this project, we try to perform an age classification by decades firstly in the Catalan CommonVoice Dataset and then add the Spanish Dataset and English Dataset to have more data. To reach our purpose Deep Learning techniques are used to implement the classifier. The most common backbones are used such as Resnet and VGG. Furthermore, we use an attention encoder to encode the Mel-Spectrogram features. In contrast to statistical pooling methods like average pooling, Attention Pooling layers and various Attention Mechanisms are used in all backbones to perform pooling and reduce the dimensionality of the feature vector derived from the Front-End architecture. In this study, we will compare two different models, the first with an AM-Softmax in the final layer and the other with an AM-Softmax combined with Ordinal Regression.