Data Collection and Development of Bengali ASR and TTS for Conversational AI-based Automated Advisories in the Agriculture domain

Abstract

This paper presents an indigenous work of text and speech data collection and organization in the agriculture domain and developing a structured Agriculture Knowledge Repository (AKR), Automatic Speech Recognition (ASR), and Text-To-Speech synthesis (TTS) systems in Bengali language, applicable to agriculture-related automated advisory systems. Authors have searched available data sources, interacted with real farmers and experts at local farming fields and agriculture institutes, collected feedback on actual advisory requirements, and designed an initial text corpus which is then used to prepare 10,000 numbers of unique query-answer pairs with 5000 agri-keywords in the final AKR. The DNN-HMM based ASR is trained by merging general domain speech data with 40.5 hours of agriculture data newly collected using a mobile app, web-based and Interactive Voice Response (IVR) based data collection setups. TTS is developed in one male and one female voice with 20 hours of studio speech data using DNN-based architecture. Both the ASR and TTS are tested with end users in real environments and are having encouraging results. Corpus collection and system development methodology is language invariant and developed sub-systems can be readily used for web or mobile chatbot-based and IVR-based automated advisory applications in the agriculture domain.