16th Annual Conference of the International Speech Communication Association

Dresden, Germany
September 6-10, 2015

Sparse Coding Based Features for Speech Units Classification

Pulkit Sharma, Vinayak Abrol, A. D. Dileep, Anil Kumar Sao

IIT Mandi, India

In this paper a sparse representation based feature is proposed for the tasks in speech recognition. Dictionary plays an important role in order to get a good sparse representation. Therefore instead of using a single over complete dictionary, multiple signal adaptive dictionaries are used. A novel principal component analysis (PCA) based method is proposed to learn multiple dictionaries for each speech unit. For a given speech frame, first minimum distance criterion is employed to select appropriate dictionary and then a sparse solver is used to compute sparse feature for acoustic modeling. Experiments are performed using different datasets, which shows the proposed feature outperforms the existing features in recognition of isolated utterances.

Full Paper

Bibliographic reference.  Sharma, Pulkit / Abrol, Vinayak / Dileep, A. D. / Sao, Anil Kumar (2015): "Sparse coding based features for speech units classification", In INTERSPEECH-2015, 712-715.