Robust audio signal processing apparatus and method for noise

In one embodiment, each spectrometer voice and audio signal g converted to an image, and, using the mask matrix from the spectrogram image and calculating the local gradient, and divided into blocks of the size of the local gradient predetermined divided block generating a weighted histogram, and generating the connection of the divided blocks each weight histogram to generate the audio feature vector, and by a discrete cosine transform (DCT) to a set of feature sets of the audio feature vector conversion feature set, and It relates to an audio signal processing apparatus and method for generating an optimized feature set by reducing the size by removing the unnecessary area in the transformed set of characteristics.