MUMBAI, India, Nov. 2 -- Intellectual Property India has published a patent application (202641124836 A) filed by Manisha K M; and Manasa S M on October 19, 2026, for Emotion-Controlled Text-To-Speech Synthesis System And Method Using Deep Learning.
Inventors include Manisha K M; and Manasa S M.
The application for the patent was published on October 30, 2026, under issue no. 44/2026.
Abstract: An emotion-controlled text-to-speech system and method are disclosed. Input text is cleaned, tokenised and converted into numerical vectors, from which a text encoder generates a text embedding. An emotion encoder generates an emotion embedding for a user-selected emotion such as neutral, happy, angry, sad or surprise. The two embeddings are concatenated into a unified text-emotion vector that is aligned, through an attention mechanism, with the states of an autoregressive audio decoder. The decoder predicts mel-spectrogram frames, and a waveform reconstruction module, preferably using the Griffin-Lim algorithm, converts the mel-spectrogram into a speech waveform. A single trained model thereby renders the same text in different emotions. A warm-up training stage on synthetic data stabilises training, and a web interface allows text or document input, emotion selection, playback and download of the generated speech.
Disclaimer: Curated by HT Syndication.
