Which One is Dominant for Neural Network Based Speech Recognition -- Δ or Δ Δ Articulatory Parameters?

This paper presents a method that describes the effect of articulatory velocity coefficient (Δ) on neural network based speech recognition. The method consists of three stages: a) two multilayer neural networks (MLNs), where second MLN takes Δ articulatory parameters as input b) Inhibition/Enhancement (In/En) network and c) Gram-Schmidt orthogonalization before connecting with a hidden Markov model (HMM) based classifier. From the experiments on Japanese Newspaper Article Sentences (JNAS), it is shown that velocity coefficient has more effect on the phoneme recognition performance. Moreover, the proposed phoneme recognizer with articulatory velocity coefficient provides higher phoneme accuracy with fewer mixture components in HMMs.