A novel connectionist system for unconstrained handwriting recognition is proposed, using a bidirectional long short-term memory (BLSTM) recurrent neural network (RNN) with connectionist temporal classification (CTC) for sequence labelling. This approach addresses the challenges of segmenting cursive or overlapping characters and exploiting surrounding context, which have limited the performance of traditional hidden Markov models (HMMs). The system achieves word recognition accuracies of 79.7% on online data and 74.1% on offline data, significantly outperforming a state-of-the-art HMM-based system. The network is robust to varying lexicon sizes and effectively uses context for recognition. The paper compares the proposed system with HMMs, highlighting the advantages of the RNN-based approach in handling long-range dependencies and contextual information. Experiments on two large handwriting databases show that the RNN-based system outperforms HMMs, particularly for large vocabularies. The system uses a BLSTM architecture with CTC for sequence labelling, allowing direct mapping from input sequences to output labels without presegmented data. The network is trained with gradient descent and uses a language model to improve word recognition. The results demonstrate the effectiveness of the RNN-based approach in unconstrained handwriting recognition, with the system maintaining performance across different dictionary sizes. The paper also discusses the preprocessing steps for online and offline data, including normalization, feature extraction, and the use of dictionaries and language models for recognition. The proposed system shows significant improvements in accuracy and robustness compared to traditional HMM-based methods.A novel connectionist system for unconstrained handwriting recognition is proposed, using a bidirectional long short-term memory (BLSTM) recurrent neural network (RNN) with connectionist temporal classification (CTC) for sequence labelling. This approach addresses the challenges of segmenting cursive or overlapping characters and exploiting surrounding context, which have limited the performance of traditional hidden Markov models (HMMs). The system achieves word recognition accuracies of 79.7% on online data and 74.1% on offline data, significantly outperforming a state-of-the-art HMM-based system. The network is robust to varying lexicon sizes and effectively uses context for recognition. The paper compares the proposed system with HMMs, highlighting the advantages of the RNN-based approach in handling long-range dependencies and contextual information. Experiments on two large handwriting databases show that the RNN-based system outperforms HMMs, particularly for large vocabularies. The system uses a BLSTM architecture with CTC for sequence labelling, allowing direct mapping from input sequences to output labels without presegmented data. The network is trained with gradient descent and uses a language model to improve word recognition. The results demonstrate the effectiveness of the RNN-based approach in unconstrained handwriting recognition, with the system maintaining performance across different dictionary sizes. The paper also discusses the preprocessing steps for online and offline data, including normalization, feature extraction, and the use of dictionaries and language models for recognition. The proposed system shows significant improvements in accuracy and robustness compared to traditional HMM-based methods.