Speech Emotion Recognition Using MFCC Features and LSTM-Based Deep Learning Model

A team of researchers from several African universities has introduced a speech emotion recognition (SER) system that leverages Mel-Frequency Cepstral Coefficients (MFCCs) for feature extraction and a Long Short-Term Memory (LSTM) neural network for classification. The work, published on arXiv, addresses the challenge of detecting human emotional states from speech—a task complicated by speaker variability, recording conditions, and acoustic similarities between emotions like anger and excitement.

The team first processed audio from the Toronto Emotional Speech Set (TESS), converting raw signals into MFCC features that capture time-varying spectral characteristics. These features were fed into an LSTM model, which excels at learning long-term dependencies in sequential data like speech. The LSTM achieved 99% accuracy across multiple emotion classes, slightly beating a Support Vector Machine (SVM) with an RBF kernel baseline that reached 98%. The authors highlight applications in virtual assistants and mental health surveillance, where real-time emotion detection could improve human-computer interaction and patient monitoring.