Recently, the severity level of non-communicable and communicable respiratory diseases has been spreading in our society due to the rapid degradation of ecological imbalance and massive changes in individual lifestyles. Public health depends on the early detection and precise forecasting of respiratory illnesses, both communicable and non-communicable, socioeconomic development, and personal well-being. Since lung sound signal-based features are more discriminative in depicting abnormalities in respiratory cycles caused by respiratory diseases, examining lung sound signals is a highly cost-effective and non-invasive approach for early diagnosis of respiratory diseases. A novel framework based on lung sound signals using a lightweight deep learning technique is proposed to solve the early diagnosis of respiratory diseases like asthma, COPD, pneumonia, bronchiotis, etc. The proposed framework performs preprocessing of collected lung sound signal datasets to remove noises and other artifacts using filtering techniques. The extracted features are Mel frequency cepstral coefficients (MFCCs), Mel spectrograms, and other statistical features to train the built model to classify the respiratory abnormalities (e.g., eight respiratory classes, including healthy) using lightweight deep convolutional neural network (DCNN). The performance of the suggested framework is evaluated against existing methods using various benchmark settings. The proposed framework yields 98.40% accuracy in classifying seven respiratory diseases.