Accent Morphing as a Technique to Improve the Intelligibility of Foreign-accented Speech

Accent morphing aims to modify the accent of a speaker whilst maintaining speaker identity. A textindependent approach could be based on voice conversion systems which manipulate speaker identity through spectral mapping. However, it is not clear to what extent accent changes can be captured with spectral mapping alone. In this paper we implement and evaluate a text-dependent accent morphing system capable of manipulating both spectral and prosodic features. We show how accent morphing can significantly improve the intelligibility of English-accented Japanese sentences to native Japanese listeners (from 57% to 84% words correct). Analysis of the different processing conditions shows that much of the benefit of morphing comes from integrated changes to both spectrum and prosody. This suggests that text-independent morphing is unlikely to provide anything but a small increase in intelligibility.