A two-dimensional robust algorithm based on Mel-bank log-spectrum

In this paper, a novel technique named two-dimensional robust algorithm based on Mel-bank log-spectrum is presented which combines sub-band average log-spectrum enhancement algorithm and peak isolation algorithm in log-spectrum domain. Information on the long envelopes of time trajectories of log-spectral energies and robustness of the log-spectral peaks in frequency bands are used together to achieve more excellent performance in this algorithm. Comparative experiments on the Aurora 2 databases are performed between these techniques and the proposed two-dimensional algorithm. It is shown that this combined two-dimensional algorithm results in better recognition performance than each technique. Compared to the Aurora 2.0 baseline system, the word error rate reduction was 21.23% in the clean training case for this combined algorithm.