Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks

This paper introduces Connectionist Temporal Classification (CTC), a method for training recurrent neural networks to label unsegmented sequence data. CTC allows the network to learn alignments between the input sequence and the target sequence without requiring pre-segmented data or explicit alignment information. This is achieved by introducing a 'blank' label and defining a loss function that sums over all possible alignments. The method is particularly useful for tasks like speech recognition, where the timing of speech events is variable and difficult to pre-segment.