In order to exploit the full dynamic range of communications
and recording equipment, and to minimise the effects of noise
and interference, input gain to a recording device is typically
set as high as possible. This often leads to the signal exceeding
the input limit of the equipment resulting in clipping. Com-
munications devices typically rely on codecs such as GSM
06
.
10
to compress voice signals into lower bitrates. Although
detecting clipping in a hard-clipped speech signal is straight-
forward due to the characteristic flattening of the peaks of the
waveform, this is not the case for speech that has subsequently
passed through a codec. We describe a novel clipping detec-
tion algorithm based on amplitude histogram analysis and least
squares residuals which can estimate the clipped samples and
the original signal level in speech even after the clipped speech
has been perceptually coded.