Abstract
Mobile based dialogue systems in local languages provide a very suitable information delivery
channel. Many of the tasks can be addressed by designing and developing small vocabulary systems.
However, as small vocabulary systems generally try to match each input word onto one of the words in
the vocabulary, if inadvertently out of vocabulary (OOV) words are spoken, they are also mapped onto
the closed set of words in vocabulary and reduce the accuracy. The current work addresses this issue.
We present the development of mobile based dialogue system in local language (Urdu) to provide
weather information to urban and rural populations. Performance of this speaker independent automatic
speech recognition system (ASR) is evaluated by offline and online testing. In offline testing, based on
unseen dataset limited to the speakers used for training the system, 100% accuracy is achieved. In online
testing, 74.79% accuracy is achieved. Analysis shows that a significant reduction in accuracy is caused
by out-of-vocabulary words (OOV) spoken by users. Phone-based model is then added to detect and
reject OOV words and system accuracy improves to 88.24%.
Saad Irtza, Aneek Anwar, Sarmad Hussain. (2014) Improving Recognition Accuracy of Urdu Weather Service by Identifying Out-of-Vocabulary Words, Pakistan Journal of Engineering and Applied Sciences, VOLUME 15, Issue 1.
-
Views
2024 -
Downloads
172
Next Article
Article Details
Volume
Issue
Type
Language
