Biometric Authentication: A Machine Learning Approach

9.7. Concluding Remarks

This chapter has highlighted the key issues in speaker recognition and demonstrated how the neural models described in earlier chapters can be applied to speaker verification. The results demonstrate that the globally supervised learning of PDBNNs can make the FAR of all speaker models very close together and that the average FAR is very small during verification; the ad hoc approach used by the EBFNs and GMMs is not able to do so. This chapter also demonstrated that PDBNNs and GMMs are more robust than EBFNs in recognizing speakers in noisy environments. Although PDBNNs and GMMs show better generalization against additive noise at different SNR, their performance is still unacceptable at low SNR. This suggests the necessity ...