Design and development Of an enhanced speech emotion recognition Algorithm through novel adaptive fd ams

Abstract

Speech is one of the most natural ways for human newlinecommunication and conveys linguistic and speaker information. The newlinemain objective of this research is to enhance the accuracy of the human newlineemotion recognition algorithm from the speech samples. In this newlineresearch, initially, the lateral improvement of speech intelligibility is newlinecarried out by proposing a novel Fractional Delta Amplitude Modulation newlineSpectrogram (FD-AMS). A fractional concept is added along with the newlineDelta-AMS algorithm in order to remove the noise frames from the newlinesignal. The obtained features are utilized for determining the optimum newlinemask and are adapted to train a deep belief network. To extend the newlineaccuracy of intelligibility under various noise levels an adaptive newlineparameter is incorporated along with the speech enhancement process. newlineSecondly, various emotions are recognized from the speech newlinesignal. This is achieved by extracting suitable features from the noisy newlinespeech signal after the enhancement process. The frequency-based newlineparameters like tonal power ratio, spectral flux, and MKMFCC are used newlineduring the process. To recognize different emotions in the speech signal, newlinea Taylor integrated DBN classifier is proposed. Recognition accuracy is newlinevi newlinewell modeled for various emotions of special people. For achieving newlinebetter classifier response, the training is done by Moth Search Algorithm newline(MSA) along with the Standard Gradient Descent approach. Third newlinecontribution is to model an optimized tuned Taylor series with Gradient newlineDescent (GD) classifier on DBN for recognizing speech signals from newlinespecially needed ones. Here, features like Holoentropy having extended newlineLinear Prediction using autocorrelation Snapshot (HXLPS), Spectral newlineskewness, and Spectral kurtosis features are used to train the classifier. newlineFrom the analysis, the proposed fractional D-AMS method newlineacquired a higher PESQ of 2.9928 and a smaller RMSE of 0.0092. The newlinedeveloped adaptive FD-AMS produced a higher PESQ of 2.4034 and a newlinesmaller RMSE of 0.0172 at 5db and for 15db level PESQ value is 3.022, newlineRMSE value is 0.009