[QUESTION] [TTS] 'num_elements_batch_sampler' loses the randomness of the samples about espnet

[sw005320]

Good question.
The reason for this implementation is to make the balance of random shuffling and GPU memory usage.

Actually, we had an experiment before (7 years ago) for ASR between utterance-level shuffling and batch-level shuffling, and the difference was marginal (but this experiment causes different effective batch sizes, and the comparison could have been better).
Also, some people even sort it from short to long for all utterances and report that it is better (due to curriculum learning effects).
So, the entirely random shuffling may not be needed.

However, this is an old experience.
Nowadays, many technologies have changed, and we may have different conclusions. It's worth revisiting.
Also, we started to use fixed-length utterances (with padding) in some projects, where we can perform random shuffling for all utterances.

It would be great if you could do some investigations.