Figure 12

Significance levels of word error rate difference between the top four models. The histograms are created from 10,000 bootstrap samples of the test data, as described in Section 4.2. WERdiff displays the average difference in word error rate performance between the two methods. Positive values indicate that the first method has lower error than the second. The percentage values indicate the tail mass of the histogram to that point.Boost versus HMM embedE-Boost versus HMM embedBag versus HMM embedE-Boost versus BoostBag versus BoostBag versus E-Boost