MUMBAI, India, Sept. 14 -- Intellectual Property India has published a patent application (202611086526 A) filed by Noida Institute Of Engineering And Technology Niet on July 15, 2026, for Secure Multilingual Text-To-Speech System With Contrastive Grapheme-Acoustic Alignment And Voice-Cloning Prevention.
Inventors include Aditee Mattoo; and Mohit Kumar.
The application for the patent was published on September 11, 2026, under issue no. 37/2026.
Abstract: The present invention discloses a secure multilingual text-to-speech synthesis system (100) comprising a multilingual grapheme processing module (110) for character- level tokenization across languages, a contrastive grapheme-acoustic alignment engine (120) employing a frozen wav2vec 2.0 Large extractor (122) and attention-based pooling (126) to generate speaker-conditioned embeddings replacing grapheme-to-phoneme conversion, a speaker identity conditioning module (130) with a consent verification unit (134) enforcing ECDSA-signed authorization, a voice-cloning safeguard module (140) jointly training an audio watermark embedding unit (142) with a HiFi-GAN V1 vocoder (156) to embed imperceptible 128-bit provenance watermarks during waveform generation, an ECAPA-TDNN deepfake detection classifier (146), and an edge deployment module (160) utilizing model distillation and eight-bit quantization for resource-constrained hardware. The system achieves word error rates of 7.5 percent for English and 14.1 percent for Mandarin without pronunciation dictionaries, watermark detection accuracy exceeding 98 percent, and deepfake detection equal error rate below 2 percent. [Figure 1]
Disclaimer: Curated by HT Syndication.
