Meta-aprendizado para a verificação de falante com áudios de curta duração

• Ano de defesa: 2022
• Autor(a) principal: Souza, Lucas Alcântara
• Orientador(a): Soares, Anderson da Silva
• Banca de defesa: Soares, Anderson da Silva, Cândido Júnior, Arnaldo, Galvão Filho, Arlindo Rodrigues
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Goiás
• Programa de Pós-Graduação: Programa de Pós-graduação em Ciência da Computação (INF)
• Departamento: Instituto de Informática - INF (RG)
• País: Brasil
• Palavras-chave em Português: Verificação de falante; Biometria por voz; Meta-aprendizado; Curta duração; Independente do conteúdo
• Palavras-chave em Inglês: Voice biometrics; Speaker verification; Meta-learning; Short duration; Text-independent
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: http://repositorio.bc.ufg.br/tede/handle/tede/12072

In practical scenarios, a speaker verification model system must be able to identify a person given audios of any durations. However, existing speaker verification systems have low performance when dealing with short-length audios. To face this problem, the MLVL (Meta-Learning Variable-Length) approach was proposed, which consists of using audios with different durations within the same episode in the meta-learning of a prototypical network. The objective is to become text-independent speaker verification more robust to the context in which the verification audio is short-length. Models trained with the MLVL approach were evaluated in three different scenarios of short-length audios, obtaining 2.55% as the lowest EER (Equal Error Rate) value. Evaluating such models in audios with longer durations, the lowest EER value obtained was 2.40%. The results surpassed those obtained by several studies in the same scenarios, demonstrating the potential practical application of the proposed MLVL approach in a voice biometrics system.