Speaker normalisation remains a very significant problem in speech research. One of the most immediate applications for a solution would be in the area of multiple-speaker speech recognition systems. These systems are faced with the task of assigning phonetic labels to portions of input speech, a task which is extremely complicated due to the enormous amount of variability within a phonetic class. Finding a good normalisation transformation would reduce this variability. Theoretical aspects of speech related studies would also benefit from a normalisation solution as it should lead to a greater understanding of the essential acoustic correlates that define a sound. A solution would aid researchers in the areas of perception and psycholinguistics. Normalisation techniques could also contribute to speech synthesis applications, especially in the area of producing multiple voices. This paper describes a frequency domain shift function which reduces the amount of inter-speaker variance within a phonetic class. The shift function is dependent upon the speaker's geometric mean pitch. The shift function is easily parameterised in a piece-wise linear fashion. Application of the shift allows a 15.8 - 17.0% reduction of variance. This reduction falls within 0.2% of the optimal pitch-only shift function for the data studied. In addition to variance reduction and recognition applications, this shift is easily applied as a means for warping speaker quality. This technique is applicable to synthesis systems where multiple voice qualities are desired.
Bibliographic reference. Tuerk, Christine / Robinson, Tony (1993): "A new frequency shift function for reducing inter-speaker variance", In EUROSPEECH'93, 351-354.