Text-independent speaker recognition using normalization compensation transformation

Based on the acoustic characteristic of frame likelihood probability output by Gaussian Mixture Model (GMM) which was the best text-independent speaker recognition model,normalization compensation transformation as a non-llnear transform method was presented.The theory analysis and experiment showed that it could improve recognition ratio 3.7% and reduce the error recognition ratio 45.1% as compared with Maximum-Likelihood (ML) transformation.The result showed:normalization compensation transformation should be adopted for cancelling the influence of variations in speech characteristics,noise and model mismatch;Process on frame likelihood probability output by GMM is effectual way of decreasing the influence of noise and improving the recognition ratio.