This paper considers the problem of signal clipping and studies the performance of speech recognition in the presence of clipping. We present two novel algorithms for restoring the corrupted samples. Our first approach assumes that the original undistorted signal is band-limited and exploits this to iteratively reconstruct the original signal. We show that this approach has a monotonically nonincreasing mean square error. In the second procedure we model the signal as an auto-regressive process and develop an estimationmaximization (EM) algorithm for estimating the model parameters. As a by product of the EM procedure the signal is recovered. The effects of these methods on the accuracy of speech recognition are studied, and it is shown that over 4% relative word error rate reduction can be achieved on speech utterances with more than 0.5% clipped samples.
Bibliographic reference. Maymon, Shay / Marcheret, Etienne / Goel, Vaibhava (2013): "Restoration of clipped signals with application to speech recognition", In INTERSPEECH-2013, 3294-3297.