Robust speech recognition in noisy environment using perceptual features and adaptive filters

The usefulness of perceptual features are emphasized in this work, which is derived from input speeches and adaptive filtering algorithms for developing a robust speech recognition system in a real world noise environment is also proposed. NOIZEUS database is utilized to create the noisy speech. Speech recognition system is evaluated on noisy test speeches at 0 dB SNR and based on the experimental results the robustness of the proposed algorithm is revealed in recognising continuous speeches. Based on the computation of the Euclidean distance between test features and clusters with and without the use of adaptive filtering algorithms is utilized for this. The performance of these features and adaptive algorithm for noise removal is tested on speeches chosen from “NOISEUS” database with noise samples taken from “AURORA” database. These perceptual features are evaluated with an average accuracy of 53% and 100% for speech recognition for the test speeches at 0db SNR without and with the use of adaptive filtering algorithm respectively. Here an adaptive filter with eight filter coefficients is adaptively changed to generate the noise reduced speech. The perceptual features are extracted from this and subsequently applied to the speech models and based on the minimum distance criterion, correct speech is identified, and the accuracy is found to be 100%.