Users simply authorize their accounts through Dropbox and the service automatically creates a new folder for you with all of the baseline templates for both a barebones landing page and your admin interface (so you can customize that, too).
Table 3 provides the phone accuracies of the five broad phone classes (vowels, semivowels, stops, fricatives, and nasals) and the accuracy of silence for the HMM baseline and template matching.
On the Nov92 20k-word trigram WSJ task, the proposed method improved the state-of-the-art template baseline without prosodic information and led to a relative word error rate reduction of 7%.
In Table 5, we compare the recognition word accuracies between the HMM baseline and the template-based methods.
In Figure 6, we compare the phone recognition performances by using the HMM baseline and the template-matching-based lattice rescoring with the local distances of Mahalanobis, NLL, LLR, and KL divergence.
Comparison on phone accuracies (percent) from the HMM baseline and the template-matching-based lattice rescoring with the local distances of Mahalanobis, NLL, LLR, and KL, where in the last three cases 1GMM was used in labeling each frame vector.
Relative to the decoding time per frame of the HMM baseline, when all templates were used, the test per-frame labeling overhead was 40% and the rescoring overhead was 22%, and hence the overall computational overhead per frame was 62.0%.
Although in the current work we used the basic acoustic modeling techniques to train our HMM baselines, the proposed template matching methods can take advantage of and add value to more advanced GMM/HMM systems, and as such they are promising for further improving the state-of-the-art speech recognition.
Each real sweep, mock baseline sweep, and the template were smoothed with a 10-ms moving average.
