کنفرانس مهندسی برق ایران

صفحه اصلی / بیست و نهمین کنفرانس مهندسی برق ایران

Attractors Manipulation in Denoising Autoencoders for Robust Phone Recognition

نویسندگان :

Shaghayegh Reza¹ Seyyed Ali Seyyedsalehi² Seyyedeh Zohreh Seyyedsalehi³

1- دانشگاه صنعتی امیرکبیر 2- دانشگاه صنعتی امیرکبیر 3- دانشگاه صنعتی امیرکبیر

کلمات کلیدی :

deep neural networks, denoising autoencoders, attractors, acoustic landmarks, robust recognition, phone recognition

چکیده :

Autoencoder Neural Networks are capable to filter out unwanted variabilities; however, their performance will degrade if their attractors and their basins of attraction are not properly adjusted. In this paper a heuristic method is proposed to increase the number of attractors shaped in desired points and expand their basins of attraction. These well-formed attractors can compensate unwanted variabilities and hence increase the chance of robust recognition. The effectiveness of this method is shown on synthetic data and is compared with another attractor manipulation method called cyclic method. In addition, the performance of this method on phone recognition task has shown 22.1 percent relative increase in the number of attractors and 4.2 percent relative improvement in the phone error rate on Farsdat database.