Dynamic segmentation of vocal extract for Assamese Speech to Text Conversion using RNN

The current work proposes a prototype Speech to Text Conversion System (STSC) in Assamese language using Linear Predictive Coding (LPC) and Recurrent Neural Network(RNN). The LPC features are extracted from utterances of isolated phonemes of Assamese language (a major language of North-East India). These are used to train a RNN by a proposed dynamic method. The proposed method segments an utterance with an optimal dynamic criterion to improve the success scores during testing of the STCS system. The proposed method dynamically adjusts the length of the windows required for recognizing different phonemes. The performance of the proposed method is compared with a conventional static RNN based STCS system which is trained using prior knowledge about length of windows required for recognizing different phonemes.