A Convolutional Recurrent Neural Network (CRNN) Based Approach for Text Recognition and Conversion of Text To Speech in Various Indian Languages

Abstract

The field of Image Processing features a huge impact on the sector of computing and hence it can help us capture anything that we can analyze it for our own purposes. The OCR (Optical Character Recognition) provides the capability of automated identification techniques that fulfill the necessity of automation in several applications. We present a system which use CRNN (Convolution Recurrent neural network) and STN (Spatial transformer network) so as to process the image of any text. It involves several steps like pre-processing, feature extraction and text to speech conversion. Our system uses the technique for converting the textual content of any document to a computer readable format. Then our system converts this computer readable document into speech using T2S package and translates the speech into several regional Indian languages. Finally we evaluate the system in terms of accuracy, which provides 90.71% accuracy on the test dataset. On the contrary the accuracy of Google’s pytesseract on the same test dataset was only 35.4%.