The non-native canonical accent—And how to use it

Abstract

One of the key benefits of modern globalization is cross-pollination of cultures. This global community encourages diversity of thinking and communication across language boundaries. However, learning to speak a foreign language is predictably challenging and non-native speakers offer a quasi-infinite source of pronunciation variation. This variation presents a challenge to AI-based technologies where we make assumptions about user behavior and expected runtime input to the models. Inconsistent usage therefore is difficult to model successfully. A study of pronunciation patterns from a set of non-native speakers uncovered some of the underlying dynamics of different levels of non-nativeness. Anchored in key differences in phoneme inventory and phonotactic distance between L1 and L2, the notion of a non-native canonical accent is introduced. This accent captures a significant portion of the speakers in the data and allows us to describe the variation with the precision needed to model it. If we can more consistently anticipate the characteristics of the variation, we're better positioned to handle it. This means that by explicitly focusing on the pronunciation patterns that are likely to be challenging to non-native speakers, we can adapt the technology to build more robust speech recognizers and more personalized foreign language learning experiences.