MUMBAI, India, Nov. 2 -- Intellectual Property India has published a patent application (202641124063 A) filed by Koneru Lakshmaiah Education Foundation on October 16, 2026, for Wav2vec 2.0-Based Bilstm System With Automated Selection Of Optimal Pooling Strategy For Speech/Music Classification.
Inventors include Ch. V. Sai Abhiram; N. Vihas; G. Ruthika Lakshmi; D. Udayanjali; and Akhilesh Kumar Dubey.
The application for the patent was published on October 30, 2026, under issue no. 44/2026.
Abstract: A system (200) classifies audio segments as speech or music. A frozen Wav2Vec 2.0 model (104) produces contextual representations from the waveform after energy- based silence removal (102); the representations are grouped into fixed-duration segments (106) and processed by a BiLSTM network (108). Attention (110a), mean (110b), max (110c) and last-hidden-state (110d) pooling units produce alternative fixed-length segment representations, which are not fused. In a selection phase, a classifier is trained for each strategy and its performance is determined on a validation subset held out from training (112); an automatic pooling selection module (114) selects the strategy with the highest accuracy and stores its identifier (210). In classification, only the selected pooling unit is executed and a classification network (118) outputs a speech/music decision (120) for each segment, which may control downstream audio processing (214). On three public datasets, different strategies were selected, each matching or exceeding an MFCC/CNN baseline.
Disclaimer: Curated by HT Syndication.
