MUMBAI, India, July 7 -- Intellectual Property India has published a patent application (202511067039 A) filed by Deepmindsecure Technologies Private Limited on July 14, 2025, for Dialect-Conditioned Transformer Asr System With Fixed Dimensional Block-Level Embeddings.
Inventor includes Saurabh Vajpayee.
The application for the patent was published on July 03, 2026, under issue no. 27/2026.
Abstract: The present disclosure provides a method of recognising speech that includes receiving an audio signal (101) representing uttered speech, extracting acoustic feature vectors, and generating, by a dialect-embedding generator (103) including a self-supervised neural network, a fixed-dimensional dialect embedding corresponding to an administrative block of a geographical region. The method includes optionally concatenating a region meta-vector (104) encoding block code, district code and linguistic family information to the dialect embedding to form an augmented dialect vector, and conditioning both an encoder and a decoder of a Transformer neural network with the dialect embedding or the augmented dialect vector. The method includes decoding the acoustic feature vectors to produce a textual transcription (140), determining a confidence score by a confidence module (130), and when the confidence score is below a threshold, substituting the dialect embedding with a parent-dialect embedding (150) corresponding to a higher-level geographical region to produce a textual transcription (140). (FIG. 1)
Disclaimer: Curated by HT Syndication.
