Um estudo da relevância da dinâmica espectral na classificação de sons domésticos

• Ano de defesa: 2016
• Autor(a) principal: Duarte, Dami Doria Narayana
• Orientador(a): Montalvão Filho, Jugurta Rosa
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Sergipe
• Programa de Pós-Graduação: Pós-Graduação em Engenharia Elétrica
• País: Brasil
• Palavras-chave em Português: Engenharia elétrica; Som; Medição de ruído; Ondas sonoras; Espectros de Markov; Sons domésticos; GMM; MFCC; Delta MFCC; Dinâmica espectral; Monitoramento doméstico; HMM; Simbolização; PEEB
• Palavras-chave em Inglês: Domestic sounds; Spectral dynamics; Daily sound recognition; Audio surveillance
• Área do conhecimento CNPq: ENGENHARIAS::ENGENHARIA ELETRICA
• Link de acesso: https://ri.ufs.br/handle/riufs/5021

This work presents a study of the spectral dynamics characteristics of audio signals. More specifically, we aim at detecting regularities that can be modeled in typical domestic sounds, in order to classify them. Our starting point is the work of Sehili et al. [2], in which a household sounds classification system based on GMM is proposed. The Sehili system is reproduced in this work as a baseline system. Following the same protocol of experiments, a 73 % recognition rate is achieved. Afterwards, three sets of experiments are performed, arranged so that each new approach incorporates a new technique to highlight a different aspect of the spectral dynamics. The first technique is the insertion of the discrete gradient information of feature vectors, a strategy aimed at a local spectral dynamic analysis, and resultes in a perceptible increase in recognition rate. The next experiment is conducted with a HMM based classifier, in which the spectral dynamic should be encoded in state transition probability matrices. The tests with the HMM do not result in improved recognition rates. The last experiment is based on a features extraction method, proposed by the author, called Patterns of Energy Envelope per Band (PEEB). The PEEB is an extractor that highlight the signal spectral dynamics inside narrow bands. In domestic sounds recognition tests, the classification system based on a combination of PEEB, MFCC and GMM strategies resulted in a significant improvement over all other systems tested. We conclude, based on our results, that the spectral dynamics of the studied dataset plays an important role in the classification task. However, the approaches for spectral dynamic information extraction, studied in this work, are not definitive, for it is clear that they can be further developed. For example, in the case of PEEB, the recognition rate is strongly dependent on the sound class, suggesting more elaborate forms of fusion of PEEB and MFCC features for each class.