Automatic segmentation and clustering of speech using sparse coding

We investigate the application of sparse coding and dictionary learning to the discovery of sub-word units in speech. The ultimate goal is to generate pronunciation dictionaries that could be used for automatic speech recognition (ASR). A dictionary of sparse coding atoms is trained to code a subset of the TIMIT corpus. Some of the trained units exhibit strong correlation with specific reference phonemes. It is found that our sparse coding model does not place sufficient constraints for the activation of atoms to be temporally isolated, which rules out its direct application to speech segmentation. We also investigate the consistency with which orthographically identical utterances are coded. We find that the sparse coding model used in this study generates codes that contain too much variation for it to be useful for generating pronunciation dictionaries for ASR.