Modelo com base na lógica fuzzy para o reconhecimento de diferentes estados emocionais a partir da voz

Abstract

INTRODUCTION: Voice is produced by a complex neurophysiological process that involves several structures and systems of the body, and can be influenced by the individual's emotional state and personality traits. The construction of a model that uses the voice signal in different emotions will serve as a basis for the development and development of recognition systems that are robust in the automation of emotion recognition. The definition of acoustic characteristics specific to each emotional state may allow the effective separation of emotions through voice pattern recognition techniques. OBJECTIVE: To develop a model based on fuzzy logic capable of recognizing emotions from the voices of Brazilian Portuguese speakers. METHOD: This is a technological, descriptive, observational and cross-sectional study that used secondary data from the Brazilian Voice Bank in Variations of Emotions - EmoVox-BR, which has a data set composed of 182 sound signals associated with the basic emotions: joy, fear, sadness, anger, surprise, disgust and neutral state, produced by 26 professional actors and actors in training. Acoustic parameters such as fundamental frequency (fo), jitter, shimmer, glottal noise measures Glottal to Noise Excitation Ratio (GNE) and Harmonics-to-Noise Ratio (HNR), cepstral measures Cepstral Peak Prominence Smoothed (CPPS), Mel-Frequency Cepstral Coefficients (MFCC) and acoustic-prosodic parameters of fo, duration and intensity were extracted from the EmoVox-BR data. From these, the emotion recognition model was developed. RESULTS: The study presented an innovative model based on fuzzy logic for the recognition of emotional states from voice, with the integration of carefully selected acoustic and prosodic parameters. The model uses 18 input and 7 output parameters. Twenty-three fuzzy rules were implemented to discriminate seven emotional categories. The model presented an accuracy of 89.01% and a Kappa coefficient of 87.18%, with sensitivity ranging from 76.92% for happiness to 100% for sadness and neutral state. Specificities were higher than 93% in all categories, indicating a high capacity

for emotional differentiation. Overlapping variables, such as some acoustic and mel- cepstral parameters, were removed to simplify the model and refine its performance. In

addition to optimizing simplicity and efficiency in emotion recognition, the model outperformed conventional methods in robustness compared to machine learning algorithms such as Random Forest and Kernel SVM. CONCLUSION: The emotion recognition model, developed with fuzzy logic from the human voice, achieved high accuracy and precision in differentiating emotions. The result predicts promising applications in affective computing and interactive technologies; the approach demonstrates effectiveness in capturing emotional nuances, with adaptation to different contexts to make human-machine interactions more empathetic and personalized.