Comparative Study on Different Architectures used for the Automatic Speech Recognition of Malayalam- A Low Resource Language

Speech is the most natural form of human communication and is one of the most information-laid signals. Sound waves have high and multi-layered temporal-spectral differences. In ASR acoustic features are mapped to the basic units of sound called phonemes, and then reconstructed into words and sentences of a particular language. Malayalam belongs to the Dravidian family of languages. A Speech recognition programs have a broad range of applications. Malayalam is a low resource south Indian language because of lack of availability of enough speech-to-text corpora. It also faces challenges like code switching, different accents in different regions, non-native speakers etc. Development of Malayalam speech recognition system is at its beginning stage. Only a few researches were done in this field. Speech features can be extracted and modeled using different methods. The purpose of feature extraction is the reduction of the high dimensionality of audio signal without losing the quality of speech data. The paper reviews the feature extraction and modeling techniques used in Malayalam Speech Recognition systems.