LSTM Architecture Valuation in Recognizing Speech-Based Emotion

Using Recurrent Neural Networks Long Short-Term Memory (RNN-LSTM) as a classification model can yield varied results. This research explores the effectiveness of LSTM as a classification model for Speech Emotion Recognition (SER) system, and how its architecture influences accuracy. The study compared various LSTM models with different numbers of layers and cells configurations using the TESS dataset. These models were evaluated through univariate ANOVA and correlation tests. Our results indicate that the number of LSTM cells significantly impacts accuracy, while the number of layers does not. The single-layer LSTM with 128 cells achieved the best training accuracy at 0.994 and test accuracy at 0.993.