Identificación de errores frecuentes en la pronunciación del español en hablantes no nativos utilizando redes neuronales profundas

Abstract

This document presents models based on CNN and RNN-LSTM for Spanish mispronunciation detection in Anglophone and Francophone students from the United States and Canada. The results of the experimentations show an average accuracy of 88%, highlighting performances above 90% in well-pronounced words and above 70% in identifying mispronunciations. The models implement Data Augmentation, L2 Regularization, and Dropout techniques to prevent overfitting during training with the built database. Also, they implement Layer Normalization, Batch Normalization, and Data Pipeline techniques to reduce training times per experiment.