A Novel Approach to Noise Reduction in Audio Signal

Voice activity detection is often used at the beginning of speech processes, such as speech recognition, and speech compression. It is used to detect the presence of a speaking voice, and it directs the execution of the application accordingly. There are certain features that are used to detect the presence of the voice. The efficiency of the algorithm is directly related to the number of the features used. Since the classical Voice activity detection algorithms are usually developed using Short Time Energy, they are very sensitive to the low signal-to-noise ratio values, therefore they may not provide the desired results. The parts with a speech and the parts with noise were attempted to be distinguished by using real-time sounds for the features used in the solution. In this study, three features, such as Short Time Energy, Periodicity, and the Spectral Flatness, were used to detect the voice. The desired results have been obtained by using these three features, even at low SNR values. This method has been observed to achieve better performance especially at low SNR values than conventional methods