Incremental language model and dynamic decoder for Tamil chat system

The paper gives an overview about a newly designed chat system “Tamil Vayadi” that recognizes Tamil voice input from the sender, converts it to text and transfers to the other end in a serialized form. This is received by the receiver in text form and presented in a new chat window. The system tries to eliminate some of the problems that are profound to Speech processing systems. The Input speech is split into words and each word is pre-processed (Normalization and Noise Removal). This is then fed into segmentation algorithm which splits the words into phonemes and PLP values of these phonemes are extracted which is stored for comparison during recognition. During recognition the PLP features of the extracted phonemes are compared with the PLP values of the available phoneme for converting the speech to text. A Bigram Language Model and a Decoder are designed for this purpose of conversion. We have introduced and incorporated the Tri phone effect for effective detection of phonemes. A learning module is also incorporated to effectively identify the phoneme. Using these improvements, we have achieved a word identification rate of 70%.