Train a Tensorflow model with dynamic data shape

Question

I have audio data, which I calculate there Mel-Spectrogram, the audios has different duration (between two seconds and one minute), af