Towards Simple Methods of Noise Robustness

We introduce an effective and simple noise-robust feature processing technique which achieves very good results on the Aurora noisy-digits database. This technique does not require knowledge of the noise type and level. Also, it does not require any increase in modeling parameters. It performs well both on matched and mis-matched training and testing environments. In comparison to the Aurora baseline results, it improves relative performance by 45% in the case of multi-condition training and 60% in the case of clean training. The improvement is most profound in the noisiest of cases. Our feature processing technique can be easily integrated into other noise-robust feature processing schemes and noise-robust speech models to possibly yield further improvements. In other words, the simplicity of our technique suggests that it might be generally applicable.