A preliminary study about robust speech recognition for a robotics application

In this study the specific problem of robustness in automatic speech recognition under various acoustic conditions is reviewed. The currently most used techniques to solve this problem are discussed. One specific technique (RASTA-PLP) is then implemented and compared with a more conventional approach (MFCC) for the frond-end processing for the recogniser. This technique is tested in a fictitious application (with a 26-word vocabulary, finite-state grammar, 10 speakers, and added Gaussian noise) intended for use in an autonomous robotic system. Preliminary word recognition results show promising indications that the chosen technique is suitable for improving the robustness.