Investigations on inter-speaker variability in the feature space

We apply Fisher variate analysis to measure the e ectiveness of speaker normalization techniques. A trace criterion, which measures the ratio of the variations due to di erent phonemes compared to variations due to di erent speakers, serves as a rst assessment of a feature set without the need for recognition experiments. By using this measure and by recognition experiments we demonstrate that cepstral mean normalization also has a speaker normalization e ect, in addition to the well-known channel normalization e ect. Similarly vocal tract normalization (VTN) is shown to remove inter-speaker variability. For VTN we show that normalization on a per sentence basis performs better than normalization on a per speaker basis. Recognition results are given on Wallstreet Journal and Hub-4 databases.