Optimization on decoding graphs using soft margin estimation

This paper proposes a discriminative learning algorithm for improving the accuracy of continuous speech recognition systems through optimizing the language model parameters on decoding graphs. The proposed algorithm employs soft margin estimation (SME) to build an objective function for maximizing the margin between the correct transcriptions and the corresponding competing hypotheses. To this end, we adapted a discriminative training procedure based on SME, which is originally devised for optimizing acoustic models, to a different case of optimizing the parameters of language models on a decoding graph constructed using weighted finite-state transducers. Experimental results show that the proposed algorithm outperforms a baseline system based on the maximum likelihood estimation and achieves a reduction of 15.11% relative word error rate when tested on the Resource Management (RM1) database.