الفهرس 
Advanced Machine learning algorithms for improving conversational Arabic speech recognition systems performance
SummaryOnly 14 pages are availabe for public view
 |  | from | 89 |  |  |
| | | from | 89 | | |
Abstract
The need for Automatic Speech Recognition (ASR) systems is increasing across different fields. The telecommunication field is one of these fields that ASR systems are strongly needed, as they can be very helpful in getting the corresponding text of the received calls to be used in performance evaluation and the other checks of quality control (QC), so this work introduces an ASR system trained especially to be deployed and used in real contact center departments as a replacement for human agents. This work is an attempt of contribution to the research and development efforts in the field of ASR for the Arabic language focusing on the Telecommunications domain and to the best of our knowledge, this system is the first ASR system trained to be used in real contact centers.
Building ASR systems needs a lot of audio data which may not be available for most of the spoken languages, so this work presents a novel idea to overcome the lack of data problem that helped in lowering the word error rate (WER) by up to 6% relative to using the available data only.
The Arabic language is very hard and has a lot of variations in writing styles that the same word can be written in different formats and yet have the same meaning. This work proposes a solution to unify the different writing styles which lowered the WER by up to 32.2% relative to using the different writing styles.