1. Building an NLP based speech recognition technology for emergency call centers.
- Author
-
Erukala, Sudarshan, Reddy, Prabhakar, Ramesh, Oruganti, Ramesh, Nagaram, Kumar, Atul, Prabhanjan, Bonthala, and Bolukonda, Prashanth
- Subjects
- *
SPEECH perception , *ARTIFICIAL neural networks , *LANGUAGE models , *CALL centers , *GAUSSIAN mixture models , *NATURAL language processing , *AUTOMATIC speech recognition - Abstract
The approaches of automated speech identification for spoken conversations in emergencies call centres were explored and compared therefore in research. These methodology included acoustic and linguistic models, as well as labelling techniques. Currently present speech recognition algorithms perform poorly because contact centre discussion speech has special context and is spoken in loud, emotional contexts. Consequently, the primary components of speaker verification designs and acoustical training methodologies—as well such Various investigations and analyses of symmetrical information labelling methods were performed. Various variations of Deep Neural Network/Hidden Markov Model (DNN/ HMM) and Gaussian Mixture Model/Hidden Markov Model (GMM/HMM) approaches might have been implemented and tested in order to establish an efficient language framework for conversation information. Furthermore, useful conversation system language models developed Using intrinsic and extrinsic criteria, outlined Finally, when these recommended information labelling techniques with spelling correction are compared with typical labelling techniques, they dominate the other methodologies by a significant proportion. Using the investigation's findings as a guide, we found Showed the use of spelling adjustments prior to training information for a labelling approach, trigram with Kneser-Ney discounting for a language model, and DNN/HMM for an acoustic model are efficient setups for conversation voice recognition in emergency call centres. In order to be clear, this study was Done using two distinct datasets that were gathered from emergency calls: the Dialogue dataset (27 h), which comprises the speech of the call agents, and the Summary dataset (53 h), which contains spoken summaries of those conversations summarising emergency situations. Even if the remarks were taken from the Our strategies are loosely related to particular linguistic aspects despite the fact that the emergency contact centre is in the Turkic language family of Azerbaijani, which is spoken there. As a result, it is expected that the recommended ways will also work with the other languages in the same family. [ABSTRACT FROM AUTHOR]
- Published
- 2024
- Full Text
- View/download PDF