Developing Automatic Spoken Dialogue System for Afan Oromo using Deep learning

Abstract

Dialogue system is the software that communicates with a human user through natural language on a turn-to-turn basis. The main purpose of a dialogue system is to provide an interface between a user and a computer-based application. Afan Oromo is one of the Cushitic family languages in Ethiopia which has largest mother tongue speakers in East Africa next to Kiswahili. As far as our knowledge, this language did not have automatic spoken dialogue system which can be implemented in different domains. This is one of the motivations to conduct this research. Here in this study, we have developed prototype of multi domain Afan Oromo Spoken Dialogue System (AOSDS) using Deep Learning algorithms to assist visitors of public business mall. The Statistical Approach was implemented to develop the components of our AOSDS. The freely available research toolkits and software like PyDial, NLTK and Google cloud platform were uti lized for Text and Speech data preprocessing, and ASR, TTS and dialogue manager development. We have developed cloud-based speech recognizer using Audio from 16 (8 Males and 8 Females) speakers which is around One hour long. To develop language model for A/O ASR we have used more than 5000 sentences which are collected from different sources and domains. Obviously, this is very low corpus size to develop ASR which has significant effect on performance of SDS. Like other SDS systems, the performance of our AOSDS relies on the modules which are man datory for SDS development. Among these components, the error rate of Semantic decoder for proposed system was 25.5% and the error rate of the ASR was 36.5%. As a result, the word error rate of our AOSDS is 38.6%. Lack of language resources such as text and speech data, unavaila bility of toolkits and hardware resources for this study are the major challenges of this study. With all these limitations we have found the promising result for the possibility of developing Afan Oromo spoken dialogue system.