TBscreen is a passive cough classifier for tuberculosis (TB) screening using a controlled dataset. The study enrolled 149 subjects with pulmonary TB and 46 controls with other respiratory illnesses in Nairobi, collecting 33,000 passive coughs and 1,600 forced coughs. A ResNet18-based classifier using passive cough scalograms achieved a fivefold cross-validation sensitivity of 0.70 ± 0.11. The smartphone-based model showed better performance in subjects with higher bacterial load (ROC-AUC: 0.87) or lung cavities (ROC-AUC: 0.89). Passive cough features distinguished TB from non-TB subjects and were associated with bacterial burden and disease severity. The study highlights the importance of analyzing cough features and biases in model training. TBscreen, a binary classifier using RGB images of scalograms, achieved an average ROC-AUC of 0.79 and 0.82 on datasets T1 and T2, respectively. The smartphone-based model performed best, with ROC-AUC scores of 0.83 and 0.86 for T1 and T2. The model showed better performance in male subjects, those with higher GeneXpert grades, and those with cavitary findings. The study also evaluated the impact of audio frequency, sampling rate, and devices on model performance. The best performance was achieved with a frequency range of 10 Hz to 4 kHz and a sampling rate of 44.1 kHz. The smartphone-based model outperformed other devices. The model's performance was influenced by demographic and clinical factors, with better results in older age groups, HIV-positive individuals, and those with a smoking history. The study found that scalogram-based models outperformed mel-spectrogram-based models in TB classification. The model's performance was limited by the number of subjects and the need for larger datasets. The study also noted that the model was not optimized for real-world deployment due to its sensitivity to ambient noise. The Nairobi cough dataset provides a unique opportunity to evaluate cough-based screening tools with minimal ambient interference. The study supports the feasibility of using smartphones for TB screening and highlights the potential of cough detection in identifying infectious TB cases. The study's findings suggest that TBscreen can discriminate between TB and non-TB coughs, but further research is needed to validate its performance in different settings.TBscreen is a passive cough classifier for tuberculosis (TB) screening using a controlled dataset. The study enrolled 149 subjects with pulmonary TB and 46 controls with other respiratory illnesses in Nairobi, collecting 33,000 passive coughs and 1,600 forced coughs. A ResNet18-based classifier using passive cough scalograms achieved a fivefold cross-validation sensitivity of 0.70 ± 0.11. The smartphone-based model showed better performance in subjects with higher bacterial load (ROC-AUC: 0.87) or lung cavities (ROC-AUC: 0.89). Passive cough features distinguished TB from non-TB subjects and were associated with bacterial burden and disease severity. The study highlights the importance of analyzing cough features and biases in model training. TBscreen, a binary classifier using RGB images of scalograms, achieved an average ROC-AUC of 0.79 and 0.82 on datasets T1 and T2, respectively. The smartphone-based model performed best, with ROC-AUC scores of 0.83 and 0.86 for T1 and T2. The model showed better performance in male subjects, those with higher GeneXpert grades, and those with cavitary findings. The study also evaluated the impact of audio frequency, sampling rate, and devices on model performance. The best performance was achieved with a frequency range of 10 Hz to 4 kHz and a sampling rate of 44.1 kHz. The smartphone-based model outperformed other devices. The model's performance was influenced by demographic and clinical factors, with better results in older age groups, HIV-positive individuals, and those with a smoking history. The study found that scalogram-based models outperformed mel-spectrogram-based models in TB classification. The model's performance was limited by the number of subjects and the need for larger datasets. The study also noted that the model was not optimized for real-world deployment due to its sensitivity to ambient noise. The Nairobi cough dataset provides a unique opportunity to evaluate cough-based screening tools with minimal ambient interference. The study supports the feasibility of using smartphones for TB screening and highlights the potential of cough detection in identifying infectious TB cases. The study's findings suggest that TBscreen can discriminate between TB and non-TB coughs, but further research is needed to validate its performance in different settings.