MUMBAI, India, Nov. 2 -- Intellectual Property India has published a patent application (202641125489 A) filed by Rns Institute Of Technology; Smitha B A; Dharini E; Impana P Nittur; Shri Vidya V K; and Sumukh P Shetty on October 21, 2026, for Cerebralchat: A Hybrid 3d-Cnn And Bilstm Model For Imagined Speech Recognition From Eeg Signals.
Inventors include Smitha B A; Dharini E; Impana P Nittur; Shri Vidya V K; and Sumukh P Shetty.
The application for the patent was published on October 30, 2026, under issue no. 44/2026.
Abstract: Abstract Brain-computer interfaces (BCIs) are increasingly recognized as transformative technologies that enable direct communication between the human brain and external systems without muscular activity or spoken language. This capability holds immense promise for developing assistive solutions for individuals affected by severe speech or motor impairments. Within this field, imagined speech recognition has emerged as an important paradigm since it seeks to decode silently thought words from neural activity, offering a more intuitive and natural form of interaction compared to motor-based paradigms. However, decoding imagined speech from electroencephalography (EEG) signals remains a demanding task due to the inherent challenges of EEG recording. The signals are non-stationary, exhibit low signal-to- noise ratios, and vary significantly across subjects, making robust classification difficult. To overcome these challenges, this study proposes a hybrid deep learning framework that leverages both spatial and temporal aspects of EEG data for imagined speech recognition. Specifically, the model integrates 3D Convolutional Neural Networks (3D-CNNs) with Bidirectional Long Short-Term Memory (BiLSTM) networks. The 3D-CNN extracts meaningful and discriminative spatial features from EEG topographic brain maps while retaining electrode-level distributions, thereby enhancing spatial representation. In parallel, the BiLSTM captures the sequential dependencies within EEG signals, effectively learning long range temporal structures essential for decoding imagined speech. EEG recordings are transformed into spatio-temporal representations, ensuring that both topographic brain activity and dynamic fluctuations are preserved, thus enabling the architecture to exploit complementary feature modalities. Experiments conducted on benchmark datasets demonstrate that the proposed hybrid framework achieves an average decoding accuracy of 84.67%, surpassing several recent state of-the-art methods in imagined speech recognition. The results highlight the effectiveness of combining spatial and temporal learning strategies in addressing the complexity of EEG-based speech classification. By delivering improved robustness and accuracy, the proposed framework represents a step toward practical, real-time BCI systems. These findings underscore the potential of hybrid deep learning architectures to power next- generation assistive communication technologies, paving the way for reliable and accessible solutions for individuals with communication impairments.
Disclaimer: Curated by HT Syndication.
