Number pronunciation in a multilingual environment and implications for an ASR system

Abstract

The purpose of this paper is to address the challenges and describe step-by-step solutions faced when developing an automatic speech recognition system in multilingual societies. We give a brief statistical analysis of the data that have been harvested from the internet. The harvesting process operates in a multilingual environment where code-switching is the norm. We specifically focus our attention on the challenge of number normalization, pronunciation and the variations associated with it. We then develop various systems to illustrate the effects of different approaches to modelling the pronunciation of numbers.