Underwater Speech Intelligibility Improvement Between Submarine to Surface Station in Compress Sensing Framework

Inter speech communication between submarine to surface in an underwater vessel is always unintelligible. One of the major reasons is the underwater vessel-noise which distorts the speech signal profoundly. The Compressed Sensing (CS) techniques have been widely used to enhance the quality of the noisy speech signal. However, improving the speech intelligibility (SI) of the received speech signal with the on-board equipment is a challenging task and has never been attempted before. Hence in this work the improvement in the intelligibility of the noisy speech signal is achieved by modifying the CS technique by pre-processing the signal based on different features. The pre-processing scheme is based on projecting the received speech signal onto the null-space of the noise formants. The formants herein are extracted from the features such as Linear Prediction (LP) coefficients, Mel-Frequency Cepstral Coefficients (MFCC), and chirp group-delay (GD). Experimental results show that the proposed CS scheme using different features pre-processing (which maximizes the improvement factor), achieves signifi-cant intelligibility improvement over traditional CS and other methods. The improvement factor is obtained using short time objective intelligibility (STOI) metrics.