Could someone please explain what is difference between wer(ref, hypo) when I concatenate string and wer(ref_list, hypo_list)?

[wannasleepforlong]

I am trying to calculate wer between model transcriptions and ground truths but am getting too much difference in WER, sometime even surpassing 5%! The wer using list gives better wer than concatenated string. Could someone be kind enough to explain the difference and which is better to use? Thank you!

[nikvaessen]

You can try to visualize the alignment and spot why the amount of sub/del/ins decrease when you concatenate strings.

See https://jitsi.github.io/jiwer/reference/alignment/#alignment.visualize_alignment.

Generally, if the sentences are independent, you should not concatenate them.