@MASTERSTHESIS{ 2018:289341267,
 	title = {Remoção de ruídos aditivos e segmentação de palavras-chave em áudios},
 	year = {2018},
 	url = "https://tedebc.ufma.br/jspui/handle/tede/tede/2469",
 	abstract = "The presence of additive noise is one of the main problems in digital audio recognition
 systems as they make it difficult to segment the audio relevant portions and may also
 reduce classifier performance. The main objective of this work is to develop a method of
 noise removal and segmentation in digital audio files generated by the direct observation
 method. This method is where an observer records, in audio, all the actions taken by a
 given specimen, coded in bite categories. This method preprocesses the audio files in order
 to normalize them and reduce their dimensionality, after which the SEGAN neural network
 is used to remove the noise. The audio segmentation step begins with a pre-processing that
 attenuates the signal valleys and emphasizes the peaks, similar to signal normalization.
 The pre-processing is followed by the application of the valley silencing function, based
 on the standard deviation and standardized score. Segmentation is performed by using
 a mapping function that finds the start and end times of each segment, using silence
 detection and overlapping sliding windows. The noise removal tests were performed through
 a double-blind study, using questionnaires with an unipolar 5-point Likert scale and an
 audio dataset compiled by the author, in order to subjectively measure the method’s
 quality. Quality scores reached an average of 3.56 out of 5 on noise removal and an average
 of 4.14 out of 5 on overall audio quality. The segmentation tests were performed from a
 second audio dataset compiled by the author, and obtained Dice scores of 85.10% on the
 noiseless audios, 77.95% on the noisy audios, and 76.12% on the audios that had their
 noise removed through the SEGAN network. After the results are presented, a comparison
 is made between the obtained results and some related works currently present in the
 literature.",
 	publisher = {Universidade Federal do Maranhão},
 	scholl = {PROGRAMA DE PÓS-GRADUAÇÃO EM CIÊNCIA DA COMPUTAÇÃO/CCET},
 	note = {DEPARTAMENTO DE INFORMÁTICA/CCET}
}