HPC Project: CTC loss for RNN speech recognition

One of the major challenges in speech recogntion or any other field, that concerns itself with structured predictions, is the alginment of two different sequences. The training data for training an RNN is a set of utterances, consisting of audio recorded via a regular microphone and the transcription of the spoken words. This transcription may either be in terms of phonemes or characters. It is clear, that given a fixed length of input audio signal, the number of spoken words within that signal may vary. Therefore, in constructing a function that maps from audio to transcription, such as a neural network, an essential challenge is to align the audio with transitions between characters.