Aplicação de métodos de aprendizagem de máquina profunda para redução de ruído em sinais de fala

Abstract

The digitalisation and evolution of communication media have been widely discussed topics, and the related technologies are also subject to debate. Communication has become faster and more efficient with growing technological advancements, but issues such as sound distortion due to noise, echo, and reverberations surged into existence. To mitigate these disturbances, AI can play a crucial role. This work aimed to create AI mechanisms to identify and attenuate Acoustic Echo, focusing on CNNs to identify different sound disturbances. A CNN was used to automatically identify 4 types of sound signals: clean signal, clean signal + echo, clean signal + noise, and clean signal + reverberation. The clean signals were extracted from the TIMIT database of DARPA and manipulated to obtain the remaining signals with different types of disturbances. The spectrograms of each signal were calculated and used as input for 4 CNNs. The best CNN achieved an accuracy of 98%, demonstrating a great ability to differentiate the different types of signals + disturbances under controlled conditions. In Acoustic Echo Attenuation, the desired results were harder to achieve. Various approaches were used, including the use of manipulated speech signals and different expected outputs, such as reconstructed speech signals without echo and spectrograms. DNNs, CNNs, and Convolutional Autoencoders were used, but the results were unsatisfactory, with the PESQ metric, used to evaluate the perceived speech quality in audio signals, averaging 1.12 for the Autoencoders, below expectations. In summary, the Classification task was successful and serves as a basis for more complex work. The less promising results in Acoustic Echo Attenuation indicate the need for improvements and the possible use of other techniques, such as BLSTM networks, GRU, and recurrent Autoencoders.

Date of Award	23 Jul 2024
Original language	Portuguese
Awarding Institution	Universidade Católica Portuguesa
Supervisor	Pedro Miguel Rodrigues (Supervisor) & Bruno Catarino Bispo (Co-Supervisor)