Inter-Speaker Scaling of Poly-Segmental Formant Ensembles

A linear-scaling approach is described for handling inter-speaker variations. The approach is motivated (i) by the similarity commonly observed amongst the formant-patterns resulting from different speakers’ productions of the same utterance, and (ii) by the fact that there are linear-scaling properties associated with similarity. In practical terms, linear transformations of the formant-patterns amongst different speakers are sought and interpreted as a set of scaling relations; the formant patterns are obtained from an ensemble of phonetically-varying segments. Using multi-speaker formant data on Australian English “hello”, the ensemble scales are found to explain the bulk of inter-speaker differences. The approach is independent of segmental structure; it uses only linear regression as its main computational machinery.