Jacobian adaptation with improved noise reference for speaker verification

Jacobian Adaptation (JA) of the acoustic models is a fast adaptation technique that has been used in Automatic Speech Recognition (ASR) systems to adapt the models from the training to the testing noise conditions. This technique has been tested in previous works with both Mel-Frequency Cepstrum Coefficients (MFCC) and Frequency Filtering (FF) parameters and good speech recognition results were obtained. In this work we have used the JA technique in a speaker verification system. In the previous implementations of JA for speech recognition only one reference of the training noise conditions was used to adapt all the models. In speaker recognition systems the utterances of each speaker are only used to train his/her model. Therefore, the training noise reference can be improved by estimating a specific reference for each speaker because each speaker model is trained in different noise conditions. With this approach, which we call Model-dependent Noise Reference Jacobian Adaptation (MNRJA), a better noise estimation is obtained and therefore a better adaptation of the speaker models. In our speaker verification tests the MNRJA approach outperformed the conventional implementation of the JA technique.