MUMBAI, India, Sept. 28 -- Intellectual Property India has published a patent application (202641111272 A) filed by Vellore Institute Of Technology on September 14, 2026, for Audio Signal Processing Apparatus For Vocal And Non-Vocal Segmentation With Heuristic Relevance-Guided Gated Fusion.
Inventors include S Vinila Jinny; and Jonnalagadda Vaishnavi.
The application for the patent was published on September 25, 2026, under issue no. 39/2026.
Abstract: ABSTRACT AUDIO SIGNAL PROCESSING APPARATUS FOR VOCAL AND NON-VOCAL SEGMENTATION WITH HEURISTIC RELEVANCE-GUIDED GATED FUSION An audio signal processing apparatus (100) for segmenting an acoustic signal into vocal and non-vocal regions is provided. The apparatus (100) comprises an audio input interface (102) configured to receive an acoustic waveform, a working memory (104), and a processor (106) coupled to the working memory (104). The processor (106) transforms successive analysis windows of the acoustic waveform into a time-frequency representation, computes a heuristic relevance distribution over spectral descriptors of the time-frequency representation based on frequency-interaction, formant-structure, and spectral-energy terms, and deletes from the time-frequency representation resident in the working memory (104) those spectral descriptors whose relevance falls below a threshold, thereby reducing a volume of data held in the working memory (104). The processor (106) encodes the reduced time-frequency representation through a convolutional path (108) to produce a spatial embedding and through a bidirectional recurrent path (110) to produce a temporal embedding, computes a per-frame gate from the spatial embedding, the temporal embedding, and the heuristic relevance distribution, combines the spatial embedding and the temporal embedding under the gate to produce a fused representation, and emits a frame-level label sequence indicating whether each frame corresponds to a vocal region or a non-vocal region.
Disclaimer: Curated by HT Syndication.
