Background noise robustness is one of the major concerns in speaker identification systems. Extended missing feature theory (EMFT) has been applied for this application by ignoring the unreliable parts of the input feature. Proposed is the advanced missing feature theory (AMFT) to improve the identification rate of EMFT by considering the cross-terms in a feature vector. To reduce the computational complexity of AMFT, also proposed is the fast-AMFT (F-AMFT) by applying the recursive score-calculation algorithm. Compared with EMFT, F-AMFT requires around 400 times lower complexity and gives significantly better identification rate even for non-stationary background noise environments.
展开▼
