A bilinear transform approach for vocal tract length normalization

We have developed and evaluated a set of speaker normalization procedures derived by bilinear transform (BLT) to compensate for variations in vocal tract lengths of different classes of speakers. The warping factors are estimated using the average third formants and their bandwidth, leaving out the exhaustive search. The MFCC of the testing data are transformed by the warped Mel filterbanks to match the models of the training data. The effectiveness of this set of speaker normalization procedures is examined in an experimental study performed using an isolated digit database of man, woman and children comparing to other standard speaker normalization method. The results of experiments demonstrate their capacity to achieve recognition accuracy increase of 19.5% and 16.5% at the best.