Professor Patrick J. Wolfe's Harvard Home Page

Harvard University > Division of Engineering and Applied Sciences > Electrical Engineering > Patrick J. Wolfe

Interpolation of Missing Values Using A Gabor Regression Model

Patrick J. Wolfe and Simon Godsill.

Publications, Data, and Code

To accompany the paper "Interpolation of missing data values for audio signal restoration using a Gabor regression model," submitted to the 2005 IEEE International Conference on Acoustics, Speech, and Signal Processing.

Submitted version of the paper

Data in the form of Matlab .mat files

Matlab code to be posted presently

Below are some sound examples, color figures, and explanations corresponding to the data above (as used in the paper).

In these examples, we consider a scenario typical of that encountered in audio restoration applications, in which short gaps resulting from severe impulsive noise degradations occur frequently and at random intervals. Simulations were performed in which audio time series were artificially degraded as follows: 16-bit signals sampled at a rate of 44.1 kHz were first downsampled to 11.025 kHz and then corrupted by a series of gaps of random length in the range 2-4 ms, spaced randomly with a minimum separation of 5 ms. These signals were in turn processed using a Gabor regression model, via a redundancy-two tight Gabor system derived from a 256-sample Hanning window. Additional details are available in the paper.

Piano Example

Clean Piano Signal

Degraded Piano Signal (36.5% missing data; 4.45 dB SNR)

Restored Piano Signal (10.2 dB SNR Gain)

[Piano Signal Interpolation]

Trumpet Example

Clean Trumpet Signal

Degraded Trumpet Signal (37.2% missing data; 4.23 dB SNR)

Restored Trumpet Signal (5.94 dB SNR Gain)

[Trumpet Signal Interpolation]