That is due to the audio clips being in an incorrect sample rate. Sometimes, the notebook doesn't detect an incorrect sample rate and instead continues training a model with the clips slowed down to the correct sample rate. The audio clips need to be resampled to 22kHz.