Isolated Word Recognition System for Malayalam using Machine Learning

摘要

Automatic Speech Recognition (ASR) has received greater level of acceptance as it creates speech recognition by the human machine interface. This paper focuses on developing a syllable based speech recognition system for Malayalam language. The proposed system consists of three different phases such as preprocessing, segmentation and classification. The preprocessing is performed for noise reduction, DC component removal, pre-emphasis and framing. The segmentation process implemented using Syllable Segmentation Algorithm segments the word utterances into syllables, that are inturn fed into the system for feature extraction. In the feature extraction step, we have proposed a novel approach by adding energy and zero crossing, along with MFCC features. The classification is done using Artificial Neural Network and is also compared with HMM classifier. Experiments are carried out with real-time utterances of 100 words, and obtained 96.4 % accuracy in ANN, which outperformed HMM.