Incomplete spectrogram reconstruction with kalman filter for noise robust speech recognition

Speech recognition accuracy degrades in presence of additive noise especially when recognizer´s training data are clean. Several methods have been proposed to compensate effects of noise on recognition accuracy among these methods Feature- Based Missing Feature Techniques, have shown promising results in making speech recognition systems robust. In these methods first corrupted parts of speech spectrogram are identified then prior to recognition these parts are reconstructed. In this paper Kalman-Based incomplete spectrogram reconstruction is proposed. By kalman-Based reconstruction, correlations between successive frames are used to infer missed parts of spectrogram. Kalman-Based reconstruction, improves recognition accuracy in comparison to previous most efficient reconstruction algorithm, Cluster-based reconstruction, in low SNR values. Average of 10% improvements in -5 dB in presence of different noises is achieved with noisy cluster membership identification and 4.3% in -10 dB and 3.2% in -5 dB in presence of white noise in comparison to Cluster-Based method with weighted average cluster membership identification.