Sentiment Analysis using CNN for Emotion Extraction to Synthesize Natural Speech
Year: 2025
Type: conference-paper
Abstract: The sentiment analysis and text-to-speech technologies that synthesize emotionally expressive speech promote human-computer interaction. Emotion labels, such as happiness, sadness, anger, and neutrality, are given to the diversified dataset, and significant pre-processing operations like removal of noisy characters, special characters, and irrelevant data from the dataset provide the best possible input to the model. Three deep learning models are designed and trained on the cleaned dataset: Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Convolutional Neural Network (CNN). The accuracy, precision, recall, and F1-score are checked to rate their performance. Between them CNN shows better performance at its training accuracy at 99% and validation accuracy at 91%. The second is LSTM that follows with a training accuracy of 97% and validation accuracy of 89%. The RNN carries out low performance at accuracies of 78% and 73% for the trainings and validations respectively. The CNN model is then integrated into TTS technology. As a result, emotionally nuanced speech synthesis is enabled. For virtual assistants, audiobooks, conversational agents, accessibility tools, and other emotion-aware systems, the gap in communication is bridged, and user engagement is improved as speech characteristics dynamically change to convey the detected sentiments.
Source: Procedia Computer Science
Authors: Sanjukta Mohanty, Debabrata Sahoo, Soumyaranjan Das, Arup Abhinna Acharya, Namita Panda
Institutions: Odisha University of Technology and Research, KIIT University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4410258128