ASR for Low-Resource Languages Using Multilingual and Crosslingual Information

Ngoc Thang Vu

Cognitive Systems Lab, Institute for Anthropomatics, Karlsruhe Institute of Technology, Germany

Tuesday, June 4, 2013
12:30 p.m., Conference Room 5A

Abstract:

In this talk I will present our latest investigations of rapid building an ASR system for a new language without any transcribed audio data by using ASR systems from different languages in combination with unsupervised training and multilingual confidence score. Their positive research results indicate that languages share common characteristics which can be transferred between them. Based on this idea, different approaches are presented how multilingual information can be used to improve a speech recognizer in different levels for a variety of tasks. In the preprocessing part, I will present our research on initialization schemes for Multilayer Perceptrons using multilingual data which shows significant improvements for different target languages. In the area of language modeling, I will discuss the language modeling task for conversational Mandarin-English code-switching speech, including the prediction of code switches based on textual features and their integration into recurrent neural networks to predict code-switches.

Bio:

Ngoc Thang Vu is a PhD student at the Cognitive Systems Lab, Institute for Anthropomatics, Karlsruhe Institute of Technology (KIT) in Germany. He started studying in Germany in 2004 and received his Diploma degree in Computer Science in 9/2009. After his Diploma he started working as a research assistant at KIT and currently is in the last year of his PhD, supervised by Prof. Tanja Schultz. The main focus of his PhD work is the development of an ASR system for low resource languages and accents using multilingual and crosslingual information.