Swahili Conversational Ai Voicebot for Customer Support

Abstract

Many businesses are now embracing self-service customer support such as conversational AI chatbots. However, most chatbots are internet dependent as they are embedded on websites or internet-dependent apps. Additionally, the available chatbots exploit only major languages such as English rarely used in various rural settings in Africa. This situation can be blamed on the clear disparity in the amount of NLP data for under-resourced languages critical in the development of NLP and AI applications. This research primarily purposed to develop a Swahili conversational AI voicebot for customer support contributing to the survival and development of the cultural and linguistic heritage of Swahili consequently reducing the need for users to learn a new language when interacting with customer support software. The developed voicebot evidenced a WER of 14.56%. During the voicebot’s creation, the study collected and analysed NLP data (1000 pattern and response pairs and 3hours of domain-specific speech data) for the Swahili language hence reducing the current gap in the availability of Swahili data for speech and text NLP tasks. The study also identified that while voicebots can be effectively modelled using 3rd party off the shelf solutions and no-code solutions, those targeting under-resourced languages better thrive using customized solutions. This resultant voicebot model will inform and guide the design of similar such voicebots for other domains such as healthcare or insurance.