Use of multiple classifiers for speech recognition in wireless CDMA network environments

In this paper, we address the problem and the use of multiple classi ers for robust recognition over the cellular network. The idea is to provide more variability to the system to be trained, and to support this variability with more number of model parameters. The main drawback is that the model size, and the computational complexity increases linearly related to di erent call environment. To alleviate this problem we rst introduce a new measure called the average-arc-count into the decoding process. The main advantage of this new measure is that many of the multiple classi ers can be shut down during the recognition stage if the average-arc-count of individual classi er exceeds a certain threshold limit for a given utterance. Secondly, we can also build individual classi ers with less number of parameters and without degrading the overall system performance. Experimental results on English connected digit recognition task show a string error rate reduction of as much as 40% by using the multiple classi ers when compared to individual CDMA systems.