Comparative analysis of Kannada phoneme recognition using different classifiers

Information retrieval from audio and speech is very important in the present digital world. Phonetic search (phoneme level search) is an efficient technique for searching words or phrases from audio and speech recordings. In this paper, a baseline phoneme recognition system for Kannada language is developed using Deep Belief Networks (DBNs). Phonemes are segmented from broadcast/read mode Kannada speech. 16 MFCC features are extracted from each speech frame. These features are used as input to the recognizer. DBNs are relatively new area of machine learning. The learning procedure of DBN has two steps, an unsupervised pre-training stage and fine-tuning stage. The performance of DBN for recognition of 25 Kannada phonemes is compared with the conventional methods of speech recognition such as, Multi-Layer Feed Forward Neural Networks (ML-FFNNs) and Support Vector Machines (SVMs). Experimental results show that DBNs yield a high performance as compared to other techniques with Phoneme Error Rate (PER) of 23.6 %. In another experiment conducted, shows that DBN´s performance is influenced by number of hidden units in the hidden layer chosen.