Hierarchical Accent Determination and Application in a Large Scale ASR System

Ramya Viswanathan, Periyasamy Paramasivam, Jithendra Vepa


In deploying Automatic Speech Recognition Systems (ASR) on a global scale, several challenges arise for supporting a widely used language such as English. The primary one among them is to deal with a wide variety of accents. We propose a Hierarchical Accent Determination system that deals with accent variations across large geographical regions at macro level and then the variations at the sub-regions within a selected large geographical region at micro level along with taking context cues. Eight accents [GB, US, Australian, Canadian, Spanish, Korean, Indian & Chinese] are identified at macro level and accent-specific models corresponding to the identified accents are used. The accuracy of the accent identification system is around 80% with ASR as well as using context cues such as phone language and keyboard language. The deployment of the accent identification system has improved the overall accuracy of Speech Recognition system by 10% for accented speech. It is planned to expand the approach to identify accents with significant variations found at sub-regional level in India such as Hindi, Tamil, Telugu, Malayalam and Bengali.


Cite as: Viswanathan, R., Paramasivam, P., Vepa, J. (2018) Hierarchical Accent Determination and Application in a Large Scale ASR System. Proc. Interspeech 2018, 1958-1959.


@inproceedings{Viswanathan2018,
  author={Ramya Viswanathan and Periyasamy Paramasivam and Jithendra Vepa},
  title={Hierarchical Accent Determination and Application in a Large Scale ASR System},
  year=2018,
  booktitle={Proc. Interspeech 2018},
  pages={1958--1959}
}