Performance evaluation of emotion recognition algorithms in Brazilian Portuguese language audios

Emotion recognition in humans is a multidisciplinary field that involves analyzing several types of data. Computational techniques in pattern recognition and machine learning have been applied to emotion analysis using various modalities, including gestures and facial expressions (visual signals), the lexical content of spoken or written language (textual signals), and the sound of speech (acoustic signals). Acoustic analysis leverages characteristics of speech such as frequency, tone, intensity, and harmonics, which are strongly linked to emotional states. This type of acoustic analysis has numerous applications, such as examining relationships through dialogue, enhancing human-machine interaction, and detecting psychiatric disorders, among others. While the performance of audio-based emotion recognition algorithms is well explored in several languages, there is a notable gap in the literature regarding emotion recognition in audio dialogues in Portuguese. This article aims to address this gap by evaluating the performance of three algorithms that use different models to recognize discrete emotions, happiness, anger, fear, disgust, sadness, surprise, and neutral, in Brazilian Portuguese audios. The results indicate that significant advancements are still needed for effective emotion recognition in this language.Among the algorithms studied, the maximum accuracy and F1-score achieved were 0.53, and no peer-reviewed publications were found, specifically on emotion recognition in Portuguese involving multiple datasets.