Dynamic Feature Extraction for Speech Signal Based on MUSIC

This paper described a new dynamic feature extraction algorithm to improve the robustness of speech recognition under the noise conditions. We use Multiple Signal Classification (MUSIC) spectrum to instead of the traditional spectrum estimation method, and use 24 adjacent frames to extract features. It not only adequately reflected speech dynamic characteristics, but also had lower sensitivity for speech environment. We evaluated the feature in different kinds of noise (white noise, pink noise, street noise) and different signal-to-noise ratios (-5dB, 0dB, 5dB, 10dB). Recognition experiments show that the new dynamic feature has good robustness and computational efficiency compared to Mel Frequency Cepstral Coefficient (MFCC) under low signal to noise ratios, and it plays a very good foreshadowing role in the latter speech research.