Domain Specific Intent Classification of Sinhala Speech Data

Abstract

Building an open domain automatic speech recognition(ASR) system can be accomplished by converting voice into text and performing a text classification on top of the converted text. However, with the inherent challenges in the approach mentioned above, it is not the most feasible way of the deriving intent of speech queries in a specific domain. This paper proposes a domain-specific intent classification for Sinhala language utilizing a feed-forward neural network with backpropagation. For the purposes of this research, a Neural network is trained from Mel Frequency Cepstral Coefficients (MFCC) which are extracted from a Sinhala speech corpus of 10 hours and the performance of the system is evaluated using the recognition accuracy of the speech queries. Further, the proposed solution in the paper introduces the first-of-its-kind for domain-specific intent classification for Sinhala language.