e mel spectrogram and prepares it for the mode based on the `truncation` and `padding` arguments.
        Four different path are possible:
            - `truncation="fusion"` and the length of the waveform is greater than the max length: the mel spectrogram
              will be computed on the entire audio. 3 random crops and a dowsampled version of the full mel spectrogram
              are then stacked together. They will later be used for `feature_fusion`.
            - `truncation="rand_trunc"` and the length of the waveform is smaller than the max length: the audio is
              padded based on `padding`.
            - `truncation="fusion"` and the length of the waveform is smaller than the max length: the audio is padded
              based on `padding`, and is repeated `4` times.
            - `truncation="rand_trunc"` and the length of the waveform is greater than the max length: the mel
              spectrogram will be computed on a random crop of the waveform.

        r