Session:

Hybrid System Applications
Sunday, February 29, 2004, 18.15 – 18.35

Session Chair:

Fikret Gürgen

Paper Title:

Non-linear Prediction of Speech using ANFIS: Comparison with Neural Nets

Author(s):

Dr. M.H. Savoji, Professor, Electrical and Computer Engineering Faculty, Shahid Beheshti University, Iran
A. Kaboli, Electrical and Computer Engineering Faculty, Shahid Beheshti University, Iran

Abstract:

The neuro-fuzzy non-linear prediction of long segments of speech, as long as whole vowels, using ANFIS is reported in this paper and comparisons are made when neural nets are used for the same purpose. Emphasis is put on the generalization properties of the trained fuzzy inference system when both intra-vowels and inter-vowels variability are considered. The data-base used is composed of Farsi vowels whose waveforms are sampled at 11 and 22 KHz and digitized at 8 and 16 bit resolution. The effects of sampling frequency and bit resolution on the working of ANFIS are also reported. It is shown that although results are qualitatively similar to those obtained using neural nets, ANFIS has the ability to train more quickly, in just a few epochs, and is more apt to tune in a given data set. The tuning is more pronounced when the input data is of wider bandwidth.