Emotional speech synthesis in Tamil using RAD-TTS and GST embeddings

Emotional Speech Synthesis is advancing human-computer interaction, yet Tamil lags behind due to limited high-quality emotional speech datasets and comprehensive phonetic resources. This hinders the training and evaluation of robust models. Unlike English, Tamil lacks the necessary linguistic resources, complicating emotional speech synthesis. However, transfer learning offers a promising solution by leveraging pre-trained models from other languages. This study employs RAD-TTS, GSTs, and LSTM for emotion categorization, aiming to create expressive Tamil speech synthesis. This advancement is crucial for enhancing applications like virtual assistants and educational tools, thereby improving user interaction and technology accessibility. Key evaluation metrics include Mel-Cepstral Distortion, Synthesis Speed, Mean Opinion Score (MOS), and Word Error Rate, which assess the system's effectiveness, precision, and quality in generating emotionally expressive Tamil speech achieving a MOS score of 3.8 with an Mel-Cepstral distortion (MCD) of 19.67 and Root Mean Square Error (RMSE) of 0.11. This multidisciplinary approach addresses the gap in Tamil emotional speech synthesis, promising more nuanced and natural voice interactions.