Performance Analysis of Spectrogram-Based Versus Raw Waveform-Based Deep Learning Models for Smoker Detection from Cough Audio

The classification of cough sounds for smoker detection represents a challenging task in audio processing that compares different data representation methods. This study presents a performance analysis of two prominent deep learning approaches: a spectrogram-based model, the Audio Spectrogram Transformer (AST), and a raw waveform-based model, Wav2Vec2. We used 7,561 smoker and 7,561 non-smoker samples from the CODA TB DREAM Challenge dataset. Both models were trained with five-fold cross-validation and data augmentation (SpecAugment for AST; noise, pitch, and time shifts for Wav2Vec2). The raw waveform-based Wav2Vec2 model achieved the best performance, with an average accuracy of 86.5%, an F1-score of 0.862, and an Area Under the Curve (AUC) of 0.945, completing training in approximately 49 minutes per fold. In contrast, the spectrogram-based AST model reached around 76-77% accuracy and an AUC of 0.85 in approximately 78 minutes per fold. These findings indicate that the raw waveform-based approach is significantly more effective and computationally efficient than the spectrogram-based approach for this task, offering a robust method for non-invasive smoker classification through the analysis of vocal biomarkers.