MUMBAI, India, May 1 -- Intellectual Property India has published a patent application (202641050505 A) filed by Sona College Of Technology, Salem, Tamil Nadu, on April 21, for 'adaptive feature scaling system for efficient deep learning on small speech datasets.'
Inventor(s) include Kowsalyadevi K; Swathi S; Subhasini S A; Navya S; and Sriya S.
The application for the patent was published on May 1, under issue no. 18/2026.
According to the abstract released by the Intellectual Property India: "Feature scaling is critical for stable and efficient training of deep learning models, particularly when operating on small datasets where conventional scaling methods exhibit instability and overfitting. The present invention discloses a computer-implemented system and method for adaptive feature scaling of signal-derived data, especially audio features used in speech recognition tasks. The invention introduces Adaptive Min-Max Scaling, which dynamically updates scaling parameters over local data windows, and Hybrid Scaling (Fast), which combines robust statistical scaling with logarithmic transformation to manage skewed and high dynamic range features. These techniques improve numerical stability, accelerate convergence, reduce training variance, and minimize overfitting without reliance on batch-level statistics. Experimental validation on small speech datasets demonstrates superior accuracy and training efficiency compared to conventional scaling methods. The invention enhances the technical performance of deep learning systems and is applicable to small-data artificial intelligence, speech recognition, and embedded or edge computing environments."
Disclaimer: Curated by HT Syndication.
