Taylor-DBN: A new framework for speech recognition systems

摘要

Speech recognition is a rapidly emerging research area as the speech signal contains linguistic information and speaker information that can be used in applications including surveillance, authentication, and forensic field. The performance of speech recognition systems degrades expeditiously nowadays due to channel degradations, mismatches, and noise. To provide better performance of speech recognition, the Taylor-Deep Belief Network (Taylor-DBN) classifier is proposed, which is the modification of the Gradient Descent (GD) algorithm with Taylor series in the existing DBN classifier. Initially, the noise present in the speech signal is removed through the speech signal enhancement. The features, such as Holoentropy with the eXtended Linear Prediction using autocorrelation Snapshot (HXLPS), spectral kurtosis, and spectral skewness, are extracted from the enhanced speech signal, which is fed to the Taylor-DBN classifier that identifies the speech of the impaired persons. The experimentation is done using the TensorFlow speech recognition database, the real database, and the ESC-50 dataset. The accuracy, False Acceptance Rate (FAR), False Rejection Rate (FRR), and Mean Square Error (MSE) of the Taylor-DBN for TensorFlow speech recognition database are 96.95%, 3.04%, 3.04%, and 0.045, respectively, and for real database, the accuracy, FAR, FRR, and MSE are 96.67%, 3.32%, 3.32%, and 0.0499, respectively. Similarly, for the ESC-50 dataset, the accuracy, FAR, FRR, and MSE are 96.81%, 3.18%, 3.18%, and 0.047, respectively. The results imply that the Taylor-DBN provides better performance as compared to the existing conventional methods.