Multitask learning in connectionist speech recognition
Authors
Youyi Lu, Fei Lu, Siddharth Sehgal, Swati Gupta, Jingsheng Du, Chee Hong Tham, Phil Green, Vincent Wan.
Conference
Proceedings of Tenth Australian International Conference on Speech Science & Technology
Year
2004
Location
Sydney
Links
Humans do not learn new tasks in isolation: for instance speech recognition interacts with speaker recognition. Multitask Learning (MTL) attempts to take advantage of these interactions by adding additional, related tasks to the main task. In this work, we present applications of MTL to Automatic Speech Recognition in noise, incorporating speech enhancement and gender recognition as additional tasks. A Recurrent Neural Net architecture is used to integrate multiple tasks. Results show that related extra tasks can indeed produce significant improvements to performance on the main task, i.e. ASR.