An Augmentative Phonation and Articulation System using Advanced Signal Processing Techniques as an Alternative Communication Device

Abstract

Speech recognition systems utilize acoustic signals collected using a microphone. But in individuals with speech disorders such as who have undergone laryngectomy or have vocal cord paralysis, it is not possible to collect the acoustic signals. For rehabilitation of such individuals an alternate method of communication has to be devised which is independent of acoustic signals. Facial muscles of such individuals remain intact and can be thus used for speech recognition purposes. The limited research on subvocal voice recognition demands the need to develop robust methods which can aid in rehabilitation of the effected individuals. This study aims at filling the gaps left by the previous literature. The EMG signal filtration is carried out to denoise the signals using ana advanced signals processing technique called Variational mode decomposition (VMD). VMD decomposes the input signals into its sub signals in different frequency spectra. Each frequency spectrum undergoes Iterative interval thresholding (IIT) using SIFT operator. These filtered signals are then used to extract crucial information from the signals using a novel feature extraction technique that utilizes the VMD method along with singular vector decomposition. These extracted features are utilized to classify the isolated words using Random Forest classifier. The results demonstrate the superiority of this technique, achieving an accuracy of 98.6% and 92% for a vocabulary set of 70 words and 96 words respectively. The final objective of this study was to develop a novel speech activity detection method using IMU signals as an alternate to EMG used previously in the literature. The proposed activity detection algorithm is carried out in four stages. The algorithm provides significantly better results as compared to EMG based activity detection method. The mean activation error rate (AER) for IMU based algorithm was 0.138 and for EMG based activity detection method was 0.275. These findings demonstrate the superiority of the proposed feature extraction and speech activity detection methods over alternative techniques. In conclusion, subvocal voice recognition is vital for rehabilitation, enabling silent communication and enhancing independence. Limited research and lack of IMU utilization pose challenges, but the thesis proposes a novel approach integrating spatiotemporal feature extraction and IMU data for improved accuracy and robustness. Results demonstrate the superiority of the proposed methods.