Efficient Joint Compensation of Speech for the Effects of Additive Noise and Linear System

  • Alex Acero

Proc. of the Sixth ARPA Workshop on Human Language Technology |

Published by Morgan Kaufmann Publishers

As automatic speech recognition systems are finding their way into practical applications it is becoming increasingly clear that they must be able to accommodate a variety of acoustical environments. This paper describes two algorithms that provide robustness for automatic speech recognition systems in a fashion that is suitable for real-time environmental normalization for workstations of moderate size. The first algorithm is a modiciation of the previously-described SDCN and FCDCN algorithms, except that unlike these algorithms it provides computationally-efficient environmental normalization without prior knowledge of the acoustical characteristics of the environment in which the system will be operated. The second algorithm is a modification of the more complex CDCN algorithm that enables it to perform environmental compensation in better than real time. We compare the recognition accuracy, computational complexity, and amount of training data needed to adapt to new acoustical environments using these algorithms with several different types of headset-mounted and desktop microphones.