Skip to Main Content

To simulate the conditions of online applications, which process speech as it is produced by the user, we consider a subset of features that may typically be extracted from the speech signal only up to the IPU containing the target ACW. These features are marked in Tables 4,567 through 8 with letter . With these features, we train and evaluate an SVM classifier for the three tasks described previously. Table 13 shows the results, comparing the performance of each classifier to that of the models trained on the full feature set, which simulate the conditions of off-line applications. In all three cases the on-line model performs significantly worse than its offline correlate, but also significantly better than the baseline (Wilcoxon, p < 0.05).

Table 13

Error rate of the SVM classifier on online and offline tasks.


All Functions
Disc. Boundary
Acknowledgment
Feature Set
Online
Offline
Online
Offline
Online
Offline
LXDSTMACPH (Full model) 17.4% 14.3% 10.1% 6.9% 6.7% 4.5% 
LXDS (Text-based) 21.4% 16.8% 13.5% 9.1% 10.0% 5.9% 
Word-based baseline 27.7% 18.6% 15.3% 

All Functions
Disc. Boundary
Acknowledgment
Feature Set
Online
Offline
Online
Offline
Online
Offline
LXDSTMACPH (Full model) 17.4% 14.3% 10.1% 6.9% 6.7% 4.5% 
LXDS (Text-based) 21.4% 16.8% 13.5% 9.1% 10.0% 5.9% 
Word-based baseline 27.7% 18.6% 15.3% 

Close Modal

or Create an Account

Close Modal
Close Modal