An English Pronunciation Error Detection System Based on Improved Random Forest

Abstract

The existing English pronunciation error detection methods are more oriented to the detection of wrong pronunciation, and lack of targeted improvement suggestions for pronunciation errors. With the aim of solving this problem, the paper proposes an English pronunciation error detection system based on improved random forest. Firstly, a speech corpus is constructed along with the evaluation of the acoustic features. Then an improved random forest detection algorithm is designed. The algorithm inputs rare mispronunciation data into a GAN neural network to generate new class samples and improve the uneven distribution of mispronunciation data in the sample set. The distribution rules of the pronunciation data are extracted layer by layer by stacking deep SDAEs, and the coefficient penalties and reconstruction errors of each coding layer are combined to identify the features associated with the wrong pronunciation in the high-dimensional data. Furthermore, a forest decision tree is constructed using the reduced-dimensional feature-based data to improve the pronunciation detection accuracy. Finally, the extracted 39 Mel Frequency Cepstral Coefficient (MFCC) acoustic features are used as the input of the improved random forest classifier to construct a classification error detection model. The experimental results indicate that the designed system achieves a high accuracy of English pronunciation detection.