Support Vector Machines for Segmental Minimum Bayes Risk Decoding

Segmental Minimum Bayes Risk (SMBR) Decoding is an approach whereby we use a decoding criterion that is closely matched to the evaluation criterion (Word Error Rate) for speech recognition. This involves the refinement of the search space into manageable confusion sets (ie, smaller sets of confusable words). We propose using Support Vector Machines (SVMs) as a discriminative model in the refined search space. The hope is we will be able to use SVMs effectively when the search problem is broken down into sequence of independent and simpler problems. Our first approach will be to use SVMs to make hard decisions, ie, the SVMs will output a word for each confusion set. We will then show that on using a simple voting scheme we improve upon the baseline significantly (10% relative at 9%WER) on a small vocabulary task.