An LSTM network-based real-time noise suppression tool for conducting audio and video conferences
A. Gosain, A. Sood · IET conference proceedings. · 2022
Background: A real-time noise suppression tool is an essential requirement of today's time. By using such a tool, audio and video conferences can be conducted noiselessly, even in the absence of a quiet office environment. The tool's methodology includes a signal pre-processing phase and a network computations phase followed by statistical analysis. In the pre-processing phase, various signal transformation methods are analyzed comparatively. Then, the most suitable one is applied to the input audio signal to extract its main frequencies. Afterward, these extracted features are used for training the deep neural network. Results: Many sample audio signals are fed to this tool, and it gives their cleaned versions with suppressed background noise. Statistical analysis for particular sample audio shows the amplitude difference between the two speech signals graphically. Conclusion: It is found that the Short-time Fourier Transform method is better amongst all other methods for extracting signal features due to its lossless nature. The backbone of the Long Short-term Memory Network facilitates this tool to retain any extended sequential information. An overall T statistic value of -94.920 for sample audio 1 and -208.843 for sample audio 2 solidifies the confidence in the accuracy of this tool.