Please use this identifier to cite or link to this item: http://hdl.handle.net/10609/97227
Title: Reducción de ruido en señales de audio basada en una red neuronal convolucional
Author: López Mora, Adrián
Tutor: Meler Corretjé, Lourdes
Others: García-Solórzano, David  
Abstract: This project describes a speech enhancement system implementation based on a Convolutional Neural Network (CNN). A feature transform module computes the STFT and extracts spectral phase and magnitude from the speech signal. The CNN maps the spectrum magnitude of an input noisy speech signal to an output enhanced spectrum. A reconstruction module computes inverse STFT to recover the speech enhanced audio signal. Mozilla Common Voice database, in its Catalan corpus version, is used to perform training and testing. Noisy audio samples are obtained adding AWGN with 0 dB SNR to clean speech signals. PESQ and STOI objective metrics are used to measure system performance. System evaluation shows positive results when using SNR levels as in training, while overall intelligibility deteriorates when using higher SNR levels due to phase distortion.
Keywords: speech enhancement
audio
CNN
Document type: info:eu-repo/semantics/bachelorThesis
Issue Date: Jun-2019
Publication license: http://creativecommons.org/licenses/by-nc-nd/3.0/es/  
Appears in Collections:Bachelor thesis, research projects, etc.

Files in This Item:
File Description SizeFormat 

alopezmoraPresentaciónTFG0619.mp4

Presentación del TFG26,12 MBMP4View/Open
alopezmoraCódigoTFG0619.zipCódigo implementado268,23 kBZIPView/Open
alopezmoraTFG0619memoria.pdf1,35 MBAdobe PDFThumbnail
View/Open