Text To Speech with Emotion Control Using Deep Learning

Recent advances in neural Text-to-Speech (TTS) systems have produced speech with high intelligibility and naturalness, yet most deployed systems still sound emotionally neutral. This lack of affective expressiveness limits user engagement and degrades the quality of human– computer interaction, especially in applications such as virtual assistants, audiobooks, education, and accessibility technologies. This work proposes an emotion-infused TTS framework that extends a Tacotron-based sequence-to-sequence architecture with explicit emotion conditioning. The system leverages the Emotional Speech Database (ESD) to model five emotional categories—neutral, happy, angry, sad, and surprise—and incorporates emotion vectors alongside text embeddings in the encoder–decoder pipeline. Mel-spectrograms predicted by the model are converted to waveforms using the Griffin–Lim algorithm. Experimental training on English-emotion subsets of ESD demonstrates stable convergence of mel-spectrogram reconstruction loss and the capability to synthesize perceptually distinct emotional speech, as observed through qualitative waveform and spectrogram analysis. A web-based interface is further developed to enable end-user interaction, allowing text input or file upload with selectable emotional style. The proposed system shows that explicit emotion conditioning can significantly enhance expressiveness of neural TTS without sacrificing intelligibility, and it provides a practical foundation for emotionally aware human–machine communication.