Speech Command Classification System for Sinhala Language based on Automatic Speech Recognition

Abstract

Conversational Artificial Intelligence is revolutionizing the world with its power of converting the conventional computer to a human-like-computer. Exploiting the speaker’s intention is one of the major aspects in the field of conversational Artificial Intelligence. A significant challenge that hinders the effectiveness of identifying the speaker’s intention is the lack of language resources. To address this issue, we present a domain-specific speech command classification system for Sinhala, a low-resourced language. It accomplishes intent detection for the spoken Sinhala language using Automatic Speech Recognition and Natural Language Understanding. The proposed system can be effectively utilized in value-added applications such as Sinhala speech dialog systems. The system consists of an Automatic Speech Recognition engine to convert continuous natural human voice in Sinhala language to its textual representation and a text classifier to accurately understand the user intention. We also present a novel dataset for this task, 4.15 hours of Sinhala speech corpus in the banking domain. Our new Sinhala speech command classification system provides an accuracy of 89.7% in predicting the intent of an utterance. It outperforms the state-of-the-art direct speech-to-intent classification systems developed for the Sinhala language. Moreover, the Automatic Speech Recognition engine shows the Word Error Rate as 12.04% and the Sentence Error Rate as 21.56%. In addition, our experiments provide useful insights on speech-to-intent classification to researchers in low resource spoken language understanding.