Reducing the Codebook Search Time in G.728 Speech Coder Using Fuzzy ARTMAP Neural Networks
Mansour Sheikhan, Sahar Garoucy · 2010
Abstract: Codebook search has high computational load in code excited linear prediction (CELP) speech coders. In this paper, a fuzzy ARTMAP neural network (FAMNN) is used to determine the best index of shape codebook in ITU-T G.728 speech coding algorithm. In this way, the gain value is calculated according to this index and the best index of gain codebook is determined based on the minimum distance to each of eight gain codebook values. Empirical results show that the proposed model leads to 50.7 % reduction in codebook search time as compared to the traditional implementation of ITU-T G.728 encoder. However, the degradations in mean opinion score (MOS), perceived evaluation of speech quality (PESQ) and segmental signal to noise ratio (SNR) are not significant, as well. seg Key words: Fuzzy ARTMAP Codebook search Speech encoder INTRODUCTION input patterns and correct labels in a variety of classification problems. In general, this family of neural The high delay of conventional code excited linear networks include ART 1, ART 2 [20], ART 3 [21], prediction (CELP) algorithm degrades the communication ARTMAP [22], Fuzzy ART [23], ART-EMAP [24], quality [1]. So, a low delay-CELP (LD-CELP) algorithm was dARTMAP [25], Boosted ARTMAP [26], Fuzzy ARTVar