Signal Speech Decompression using Convolutional Denoising Audioencoders for Unsupervised Machine Learning Algorithms
Houda Abouzid, Otman Chakkor · 2019
In signal procesing, audio data compression refers to the encoding information using less bits than the original representation. This method will be identified in our theoretic approach and applied for the blind source separation (BSS) problem. In this paper, we will mix and match between two types of autoencoders which are Convolutional and Denoising autoencoders. The implementation uses Keras as a principal library of neural deep learning, in order to use this resulted signals after being analysed in blind source speech separation system. We suggest the mixture of those two types of autoencoders for unsupervised learning model that reconstructs audio mixture of speech signal based on Neural Network and deep learning. Simulation results have proven that this mixture of autoencoders can make BSS easier to study and yield more performance for the signals to be included in the automated system. The advantage of this work is the originality of the solution, given an accuracy of 81% applied on real speech signals.