A study on average voice model training using vocal tract length normalization

摘要

This paper describes a new training technique of average voice model using speaker normalization. In the proposed technique, to reduce the influence of variability of formant frequency due to the difference in vocal tract lengths of speakers, we incorporate a vocal length normalization technique into training stage of the spectral part of average voice model. It is observed that average voice model trained using the proposed technique generates sharper spectral envelope than the conventional average model. Moreover, from the results of subjective tests, it is shown that the proposed technique generates average voice of better quality than models trained using the conventional or speaker adaptive training techniques.