Performance evaluation and enhancement of lung sound recognition system in two real noisy environments

This study investigates the problems associated with lung sound recognition under noisy conditions. Firstly, the effects of noise on the lung sound feature representation and the classification performance are analyzed. Two kinds of feature representations, autoregressive and mel-frequency cepstral coefficients, are used to characterize the lung sound signals. Dynamic time warping is used to categorize the lung sounds to one of the three: normal, wheezes, or crackles. Our experimental results indicate that additive noise produces a mismatch between training and recognition environments and deteriorates the classification performance with a decrease in the SNR levels. In order to compensate the degrading effect of noise on the lung sound recognition, a dual sensor spectral subtraction algorithm is applied to the lung sound signals before the extraction of lung sound features. It is observed that the proposed algorithm is capable of providing adequate performance in terms of noise suppression and lung sound signal enhancement.