Interactive Voice Response (IVR) systems have been helping businesses in providing support to their customers over a telephone network. Modern IVRs not only have speech recognition facility, but also have an AI-powered conversational system with intent classifier. Such IVRs improve user experience, but building such a system is resource-intensive and labor-intensive process. Resources are required to fine-tune and train the intent classifier every time an intent is added. The labor-intensive part of the process is creating a domain specific labelled dataset for training the intent classifier. In this paper, we propose a smart IVR system with intent classifier based on dual sentence encoders. The intent classifier used not only requires very less resources for training, but also works better than models like fine-tuned BERT in few-shot setups (when examples per intent are less). Cost of building conversational IVR is brought down even more when such intent classifier is built in conjunction with open-source technologies like Asterik (for voice server) and Festival (for text-to-speech conversion).
Article Details
Unique Paper ID: 151741
Publication Volume & Issue: Volume 8, Issue 1
Page(s): 719 - 722
Article Preview & Download
Share This Article
Conference Alert
NCSST-2021
AICTE Sponsored National Conference on Smart Systems and Technologies
Last Date: 25th November 2021
SWEC- Management
LATEST INNOVATION’S AND FUTURE TRENDS IN MANAGEMENT