Emerging Technologies in Computing. Second International Conference, iCETiC 2019, London, UK, August 19–20, 2019, Proceedings

Research Article

Automatic Speech Recognition in Taxi Call Service Systems

  • @INPROCEEDINGS{10.1007/978-3-030-23943-5_18,
        author={Samir Rustamov and Natavan Akhundova and Alakbar Valizada},
        title={Automatic Speech Recognition in Taxi Call Service Systems},
        proceedings={Emerging Technologies in Computing. Second International Conference, iCETiC 2019, London, UK, August 19--20, 2019, Proceedings},
        proceedings_a={ICETIC},
        year={2019},
        month={7},
        keywords={Speech recognition Kaldi CMUSphinx n-gram Taxi call service Speech features},
        doi={10.1007/978-3-030-23943-5_18}
    }
    
  • Samir Rustamov
    Natavan Akhundova
    Alakbar Valizada
    Year: 2019
    Automatic Speech Recognition in Taxi Call Service Systems
    ICETIC
    Springer
    DOI: 10.1007/978-3-030-23943-5_18
Samir Rustamov,*, Natavan Akhundova1,*, Alakbar Valizada1,*
  • 1: ATL Tech
*Contact email: srustamov@ada.edu.az, natavan.akhundova@atltech.az, alakbar.valizada@atltech.az

Abstract

In this research, the application of automatic speech recognition system in taxi call services is investigated. In comparison with traditional query handling systems such as live agents, Interactive Voice Response systems, type-base websites and mobile applications, the newest trend of artificial intelligence - speech recognition can be applied to make conversations in more natural way. For developing, training and testing of the system, Kaldi and CMUSphinx open-source speech recognition tools were utilized. Approximately 4 h of speech data in Azerbaijani have been processed for both tools. Testing has been accomplished in two ways; one of which is recognizing dataset from unknown speakers, and the other one is recognizing shuffled dataset. During these tests, variance and speed were investigated, along with accuracy. Kaldi showed accuracy between 97.3 and 99.6 with variance changing between 0.03 and 4.8. On the other hand, CMUSphinx attained accuracy between 95.6 and 97.8 with variance values of 0.2 and 3.8 in relatively less training time. Accomplished results were compared and used to define appropriate parameters for investigated models.