Combination of latent semantic analysis based language models for meeting recognition

Latent Semantic Analysis (LSA) defines a semantic similarity space using a training corpus. This semantic similarity can be used for dealing with long distance dependencies, which are an inherent problem for traditional wordbased n-gram models. Since LSA models adapt dynamically to topics, and meetings have clear topics, we conjecture that these models can improve speech recognition accuracy on meetings. This paper presents perplexity and word error rate results for LSA models for meetings. We present results for models trained on a variety of corpora including meeting data and background domain data, and for combinations of multiple LSA models together with a word-based n-gram model. We show that the meeting and background LSA models can improve over the baseline n-grain models in terms of perplexity and that some background LSA models can significantly improve over the n-gram models in terms of word error rate. For the combination of multiple LSA models we did however not see such an improvement.