A Threshold Denoising Algorithm Based on Mathematical Morphology for Speech Enhancement
The presence of noise in speech signals can significantly degrade the performance of speech recognition systems. A threshold denoising method based on mathematical morphology is proposed to reduce background white noise. In the method we consider speech spectrograms as images and construct binary images from a normalized 256-level gray scale spectrogram image. We take advantage of a sudden slowing in the average value (ratio of the number of ‘1’ pixels to the total pixel number) of the binary image, and use it as the threshold value to zero spectrogram elements below the threshold, normalize the spectrogram, and finally, reconstruct the original speech signal to achieve the goal of speech enhancement. The main advantage of the algorithm is fast speed that is highly desired in real-time speech processing.
This work was supported by the Natural Science Foundation of China (No. 61471111)