If you put 'chunk_size' in `inference` part of config it will be used for inference or validation instead of chunk_size from `audio` part of config, which is used for training. It can be useful for Roformer models where optimal chunk_size is different for training and inference.