Enhanced Optimization in DCNN for Conversion of Non Audible Murmur to Normal Speech based on Dirichlet Process Mixture Feature


  • T. Rajesh Kumar
  • Arumbaka Srinivasa Rao
  • K. Kalaiselvi
  • S.S. Manivannan
  • C. Shahul Hameed
  • C.M. Velu




In various applications towards mobile communication, the recognition of speech automatically plays a vital role. The user communication devices for interaction require a large amount of vocabulary recognition system, more preciseness and real time less power consuming schemas. The miniaturized battery controlled devices suffer with power consumption and large memory bandwidth. Also the speech challenge people's mobile devices require more attention. Therefore, a useful technology is proposed to convert non-audible murmur to normal speech, based on Stochastic Biogeography-based WOA (SBWOA) integrated with Dirichlet process mixture. The features like spectral skewness, Taylor AMS, spectral centroid, pitch chroma and newly developed Dirichlet Process mixture features are extracted from the input murmured speech signal and trained in the DCNN. The identification of speech is based on Deep Convolutional Neural Network (Deep CNN), which is trained by the proposed Stochastic Biogeography WOA (SBWOA). The stochastic gradient descent method, Biogeography-based optimization (BBO) and Whale optimization algorithm (WOA) are combined together to improve the results in metric analysis. The TRP, FPR and Accuracy shows the improved results of 0.9, 0.001 and 0.99 respectively.