DEEP LEARNING-BASED SPEECH RECOGNITION OF MALAY DIALECT INSTRUCTIONS USING EDGE IMPULSE

Abstract

ABSTRACT – The increasing demand for voice-controlled systems in Malaysia for assistive technology, smart home appliances, and navigation purposes is typically served with an English-only solution. This reduces the accessibility of such technologies within a non-native English-speaking community. Given Malaysia's rich linguistic diversity, this study focuses on dialect speech recognition, specifically targeting various Malay dialects. This research uses the Edge Impulse platform to present a deep learning-based methodology for recognising speech in Malay dialects. The presented system collects and preprocesses audio data using a smartphone, while the deep learning algorithms perform efficient training and classification. Our model showed that robust and dialect-inclusive Malay voice recognition is possible, with an accuracy of around 80% for the four classes of instructions tested in four different dialects of Kelantan, Terengganu, Kedah and Standard Malay. This study therefore provides more foundation towards accessibility and usability of creating more inclusive speech recognition systems which can be tuned to regional linguistic variabilities.