The presented paper describes a novel approach to detection of speech corrupted by noise. The proposed procedure is
based on fractal dimension, which is being evaluated directly from speech signal samples using two different methods:
box-counting and the approach proposed by Katz. The recordings, taken from TIMIT database, were corrupted by five
different types of noise (white, pink, hf-channel, babble and factory) with four noise amplitudes (5,10,15,20 dB).
The resulting noisy speech was the subject of the analysis. The Otsu's method was used to determine a threshold value
for differentiating between noise-only and noisy-speech segments. It has been shown that fractal dimension-based
approach provides good basis for detecting speech under a presence of noise.