RESUMO
Automatic cough detection is key to tracking the condition of patients suffering from tuberculosis. We evaluate various acoustic features for performing cough detection using deep architectures. As most previous studies have adopted features designed for speech recognition, we assess the suitability of these techniques as well as their respective extraction parameters. Short-time Fourier transform (STFT), mel-frequency cepstral coefficients (MFCC) and mel-scaled filter banks (MFB) were evaluated using deep neural networks, convolutional neural networks and long-short term models. We find experimentally that, by regarding each cough sound as a single input feature instead of multiple shorter features, better performance can be achieved. Longer analysis windows also provide enhancement in contrast to the classic 25 ms frame. Although MFCC performance is improved by sinusoidal liftering, STFT and MFB lead to better results. Using MFB and the optimum segment and frame lengths, an improvement exceeding 7% in the area under the receiver operating characteristic curve across all classifiers is achieved.