A Multiple Feature Front-end Approach to Speech in Noise

Y. Yan (USA/PRC), C. Liu, and C. Zheng (USA)


speech recognition, dynamic informationfusion.


This paper presents our recent research activities on Speech In Noise Environment (SPINE) task. Rather than focusing on finding the best single way to handle environmental noises and changes, an approach with multiple features was used. By looking at acoustic signals from different angles, it is our hope that the underlying acoustic events can be decoded better. Multiple feature front ends were used and the recognition results were combined using ROVER. A Word Error Rate (WER) of 25.6% was obtained on SPINE1 task and a WER of 36.3% was achieved on SPINE2 task. These numbers compared favorably with the best results published in literature.

Important Links:

Go Back