Detecção de edições em áudios baseada na análise tempo-frequência e em redes neurais convolucionais

• Ano de defesa: 2023
• Autor(a) principal: Cordeiro Junior, Marcos
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Curitiba; Brasil; Programa de Pós-Graduação em Engenharia Elétrica e Informática Industrial; UTFPR
• Palavras-chave em Português: Editor de audio digital; Som - Registro e reprodução - Técnicas digitais; Processamento de sinais; Aprendizado do computador; Aprendizado profundo (Aprendizado do computador); Redes neurais (Computação); Digital audio editors; Sound - Recording and reproducing - Digital techniques; Signal processing; Machine learning; Deep learning (Machine learning); Neural networks (Computer science); CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA; Engenharia Elétrica
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/33191

Tampering detection is one of the most important topics in forensic analysis of digital audio. Splicing corresponds to the insertion of a segment of signal from a different audio into the original audio record and is included among the most common categories of tampering. Convolutional neural networks (CNNs) have demonstrated effectiveness in various audio processing tasks, which motivates research into different methods of obtaining input data. The spectrogram is a useful representation for visualizing the temporal evolution of the frequency spectrum of an audio, with different signal processing techniques available for its generation. In this study, the development of an automatic splicing detection model in digital audio using CNNs was carried out. The audio spectrogram, computed using different techniques such as Short-Time Fourier Transform (STFT) on a linear scale, STFT on a mel scale, and Constant Q Transform (CQT), was directly provided to the network as input data. A comparative study was conducted to evaluate the impact of the choice of time-frequency representation on the model’s performance in correctly classifying the original and edited audios.