There is a simple telephone-based dialog system based on Asterisk + UniMRCP + Sphinx. The issue is that the system recognizes input very, very badly, despite that the accuracy for offline recognition on the same task is around 95%.
The vocabulary is small (about 20 words), possible input is defined with JSGF grammar. The acoustic models are trained on telephone speech, recorded deliberately for this task.
This works fine when we evaluate the models in batch mode. However, after plugging the models to the real system, recognition errors are enormous.
What might be the reason?
Have you tried to measure the noise of the live recognition system? Possibly a low SNR can cause bad online recognition rates. A further idea is that the silence non-silence detection of the online system is configured inappropriately.
There is a VAD indeed, and my first guess was similar to yours,
Bu the inputs to the pocketsphinx pluging have been recorded on the running system and then recognized in offline mode with acceptable accuracy...
I think the source of the problem is the cepstral mean normalization, which works in different manner for online recognition. The only question then is how to address this issue.
This could be true, and would result in different feature vectors for online and offline recognition. Online: Features without CMS and offline: features with CMS. When you have trained your model with offline feature vectors, than you cannot use them for online recognition. Thus, you have to train your models again without performing a CMS.
Well, I tried, at least as I understand it.
When training models, one can choose the CMN mode: none, prior or current.
As far as I know, prior corresponds to the live CMN update. However, using this mode for training did not help in improving the performance of the running system...
- Badges
- Science topic