Estimação cega do tempo de reverberação através de redes neurais

• Ano de defesa: 2019
• Autor(a) principal: Prates, Rodrigo Leite
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Rio de Janeiro; Brasil; Instituto Alberto Luiz Coimbra de Pós-Graduação e Pesquisa de Engenharia; Programa de Pós-Graduação em Engenharia Elétrica; UFRJ
• Palavras-chave em Português: Tempo de reverberação; Curva de decaimento; Redes neurais; CNPQ::ENGENHARIAS
• Link de acesso: http://hdl.handle.net/11422/21244

In this dissertation we investigate the performances of the main techniques for blind estimation of the reverberation time of an environment, that is, without the a priori knowledge of the impulse response of the room. These techniques model the reverberant signal to estimate the decay of the ambient impulse response in order to obtain the time that the sound takes to decay by 60 db after the interruption of the sound source, termed RT60. Among them, we highlight the method based on neural networks, which realizes the prediction of the RT60 from a stretch of the reverberant signal. This distinction is due to its ability to estimate the RT60 robustly to noise and abrupt signal variations over time. Calculations are made in the frequency domain, with the features of the training samples generated on the mel scale, before being used by the network. Results of the RT60 estimation are presented for different environments, comparing the average absolute errors obtained with different techniques. It can be concluded that the estimation errors obtained by the methods that use neural networks are similar to those obtained by the other techniques, being these errors lower in some environment configurations.Também disponível on-line.