Contribution of Difference of a Single Feature of Acoustic Models Sample Clauses

Contribution of Difference of a Single Feature of Acoustic Models. First, in order to evaluate contribution of difference of a single feature of acoustic models to achieving high confidence, we ex- amine precision values for minimal pairs which have difference in only one of those features. Features of acoustic models examined are as follows: for the decoder Xxxxxx, units in HMMs (i.e., the four types listed in Section 2.2.1), and with/without short pause state, while for the decoder SPOJUS, all the features described in Section 2.2.2, i.e., sampling frequencies, frame shift lengths, feature parameters, covariance matrices, self loop transition / du- ration control, and with/without short pause states. For each of those features, we evaluate the precision/recall of the agreement between minimal pairs and plot their precision values in descend- ing order. Figure 2 gives this plot for those with recall values above the threshold. In terms of the maximum precision values, the best performing features are “with/without short pause states” for the decoder SPOJUS and “units in HMMs” for the decoder Xxxxxx. “With/without short pause states” for the decoder Xxxxxx performs slightly worse mainly because models without short pause mod- els for the decoder Xxxxxx have word recognition rates much lower than those for the decoder SPOJUS. These results claim that the features of acoustic models most effective in achieving high confidenceare “with/without short pause Then, Figure 4 gives the precision value of a single model pair with the maximum precision value for each of the six categories. From these results, putting more emphasis on the performance against the harder speech, i.e., the broadcast news speech, we conclude to represent the degree of the contribution of those acoustic features according to the inequalities below:6