Kapre provides Keras layers that compute audio signal processing operations (STFT, ISTFT, mel-spectrogram) on GPU in real time during model execution. These layers can be inserted as the first layer of a neural network to handle audio preprocessing within the model graph, eliminating precomputation dependencies and enabling end-to-end training. It is tested against Librosa for correctness and supports both channels_first and channels_last data formats with TensorFlow 2.16+.