Missing Sample Estimation Based on High-Order Sparse Linear Prediction for Audio Signals

The restoration of click degraded audio signals is important to achieve acceptable audio quality in many old audio media. Restoration by missing sample estimation based on conventional linear prediction has been extensively researched and used; however, it is hampered by the limitations of the linear prediction model. Recently, it has been shown that high-order sparse linear prediction offers better representation of music and voiced speech over conventional linear prediction. In this paper, the use of high-order sparse linear prediction for missing sample estimation of click degraded audio signals is proposed. The paper also explores a possible computational time saving by combining the high- order sparse linear prediction coefficient determination and filtering operations. Evaluation with different types of speech and audio data show that the proposed method achieves an improvement in SNR over conventional linear prediction based filtering for all considered speech and audio data types.