Using Transfer Learning, SVM, and Ensemble Classification to Classify Baby Cries Based on Their Spectrogram Images

Babies cannot communicate with formal language and instead convey necessary messages through their cries. In babies, the first few months of their growth period are critical to the rest of their lives, as many conditions, such as deafness or brain damage from asphyxia, can be remedied if they are detected during this time period, preventing irreparable damage. The ability to differentiate between types of cries of a baby can prove extremely useful for parents with newborn children. To achieve this, we employ several machine learning, deep learning and ensemble classification techniques. In our work, we use transfer learning with the existing pre-trained convolutional neural network of ResNet50, a Support Vector Machine (SVM). We also perform ensemble classification to combine the predictions of the SVM and deep learning model to classify between different types of baby cries. Models are trained on spectrogram images of the audio files taken from the Baby Chillanto Database. We evaluate our models with ten iterations of 5-fold cross-validation and our models achieve accuracies of more than 90%.