Robust Speech Recognition

Even though the current state of art LVCSR (large vocabulary continuous speech recognition) systems performs significant better than it used to be, their performance will greatly degrade in the noisy environments. That's the motivation of Robust speech recognition

In Robust speech recognition, we can attack the problem from many different perspectives: we can study the nature of noise and try to compensate the effect of noise in speech recognition using compensation methods such as CDCN, VTS, KLT and SVD; we can identify which part of the speech has been corrupted by the noise and recover these corrupted parts using missing feature method; we can also develop different feature sets which are more robust in certain noise environment and combine these features together, this is the work of complementary feature sets and combination scheme.

กก