Joint Sparsity and marginal classification for improving Sparse Imputation performance in speech recognition

• The self-similarity nature of speech is used to improve the Sparse Imputation method.
• The similar frames of speech utterance are identified using marginal classification.
• The Joint Sparsity method is used to reconstruct noisy component of similar frames together.

Sparse Imputation (SI) is a relatively new method that reconstructs missing spectral components of noisy speech signal with the help of the sparse-based representation approaches. In this method, the redundancy of signal in the frequency domain helps to rebuild noisy spectral components from the remained reliable ones. On the other hand different parts of speech signal, despite time intervals between them, can be inherently similar to each other. In this paper, a major modification over the SI method is proposed that in addition to data redundancy property of speech signal in small regions, takes the advantages of its self-similarity nature over long intervals. By identifying mostly similar frames, using a method based on the marginal classification, the Joint Sparsity method is applied and a method dubbed as the Joint Sparse Imputation is presented. The experiments conducted on AURORA 2 data set show that the proposed method significantly improves the recognition results in different noisy conditions, compared to the original SI method.