Multi-QuadEmoNet: cat and dog emotion classification model from animal vocalization using multi-stage LSTM-GRU paradigm.

Introduction: Decoding the emotional nuances within animal vocalizations poses a significant challenge due to their intricate nature and the limitations of traditional analysis methods. Pets enrich our lives, but true companionship requires recognizing their emotional states. By deciphering signs of happiness, fear, or anxiety, humans interact to nurture a positive relationship with the pet animals. Method: This research aims to address this challenge by developing a multi-modal approach for animal emotion voice recognition system using the proposed Multi-Quadro-Emotion Network (Multi-QuadEmoNet) method. The proposed QuadEmoNet was a neural network specifically designed for handling audio data. Initially, Mel-Frequency Cepstral Coefficients (MFCCs) were used to extract features from the preprocessed audio data. The extracted features were used to train the QuadEmoNet deep learning model for animal emotion classification. Multi-QuadEmoNet contains three QuadEmoNets. The first QuadEmoNet was trained initially to predict cat/dog. Further, second and third QuadEmoNet's were implemented to classify cat and dog emotions separately after training the model for 50 epochs. Results and discussion: The proposed system utilizes a self-collected dataset containing 6,000 cat audio samples and 2000 dog audio samples. The designed QuadEmoNet model was evaluated producing highest accuracy of 95% (DMAD), and an accuracy of 90% (CMAD) was achieved for 75 and 50 epochs. Based on the proposed Multi-QuadEmoNet model, an user interface for recognizing the animal emotions was developed for predicting the animal emotions.