Entropy-Based Extraction of Useful Content from Spectrograms of Noisy Speech Signals

In the paper, we extend and apply the 2D local entropy method (2DLEM) used for signal time-frequency repre-sentation to useful content extraction from noisy speech signals. In our previous work, we presented the 2D local entropy method (2DLEM) and tested it on synthetic signals and a small number of real-world signals. We now extend the application of our method by applying it to recorded speech signals combined with noise from different sources. The database we used is commonly used in speech recognition, where tested methods usually have the best result achieved on clean signals without added noise or on denoised signals. The 2DLEM method is used for the extraction of useful content, and in this paper, we test it in real-world scenarios. Our results show promising results for all tested signals regardless of the noise source or signal-to-noise ratios (SNRs). Combining the 2DLEM method with speech recognition methods could improve the performance of speech recognition and understanding systems.