Efficient Resampling of Speech/Audio Signals in Shift-Invariant Spaces

Resampling is a necessary step at the interface of systems that require different sampling rates. Several techniques have been proposed in the literature to perform resampling. We perform a comparative study of the results of some interpolation and resampling techniques for speech and audio signals. The techniques are compared on the basis of quality of reconstruction and computational cost. We consider the following interpolation techniques: 1) linear; 2) shifted linear; 3) Keys cubic; and 4) cubic B-spline. These techniques have been extensively applied in image processing applications such as magnification and rotation. We study the efficacy of these techniques in resampling speech and audio signals. We observe that, in terms of the quality of signal reconstruction, the techniques can be ranked in decreasing order of performance as follows: cubic B-spline, Keys cubic, shifted linear, and piecewise-linear interpolation. In terms of computational cost, the interpolation techniques, arranged in increasing order of time taken are as follows: piecewise-linear interpolation, shifted linear, Keys cubic, and cubic B-spline. However, we note that the shifted linear interpolation technique gives a significant improvement in quality with just a small increase in computational cost. This might be particularly important in embedded and real-time applications where processing time is critical