Figure 6: Feature extraction:

Each raw cough recording, as shown in Figure 3, is automatically split into individual sections after which features including MFCCs (including velocity Δ and acceleration ΔΔ), linearly spaced log-filterbank energies (including velocity Δ and acceleration ΔΔ), zero crossing rate and kurtosis are extracted. For example, when using 13 MFCCs, (13 × 3 + 2) = 41 features including Δ, ΔΔ, zero crossing rate and kurtosis are extracted for each section. The number of sections, MFCCs and linearly spaced filters are used as feature extraction hyperparameters listed in Table 3.