Publication

Abstract : The proposed work utilises a combined architecture of time delay neural networks (TDNN) and multi-layered bidirectional long short-term memory (Bi-LSTM) network for the ink recognition from the handwritten text. We added a Trie beam search decoder with three smoothing algorithms such as Kneser–Ney Back off, Kneser–Ney Interpolated, and Stupid Back off to improve the performance of the model. This paper discusses the performance of the combined TDNN and Bi-LSTM network with the above-stated decoder for the ink recognition task. This paper also reports a comparison with the existing models that were implemented for this task. The analysis showed that the model with TDNN and Bi-LSTM architecture with an additional Trie beam search decoder with Kneser–Ney Interpolated smoothing algorithm using 10,000-word lexicon performed better than the model without a decoder.