Look at what the scoring function actually does. It normalizes both strings, passes them to `jiwer.process_words()`, and hands back substitutions, deletions, insertions and a reference word count [6]. Every edit costs one, so the worked example in the walkthrough, "fifteen milligrams" heard as "fifty milligrams", returns a WER of 0.2 from a single substitution against five reference words [11][19]. Read that upward and the acceptance gate gets thin: a file described as 96% accurate is a 4% error rate, which is roughly 40 wrong words in a 1,000-word transcript, and the number does not say whether they were articles or units [18].
Normalization is the part that decides whether anyone else's number means anything to you. Skip it and "3" against "three" or "Dr." against "doctor" score as errors even though the recognizer heard correctly [8]. The walkthrough treats the normalizer from the Whisper paper's Appendix C as the closest thing to a default, and describes it as stripping bracketed phrases, dropping fillers like "um" and "hmm", and expanding contractions [9]. That is also the argument the author makes for why vendor-published WER figures are not comparable [10]. For a vendor figure to transfer to your workload, you would need the same normalizer, the same reference transcription conventions, and audio that resembles yours. Two of those three are things vendors do not publish.
The entity axis is a bag-of-terms diff, not an alignment. It iterates a sorted lexicon with `re.findall` so a term said three times counts three times, then walks reference terms and removes each match from a copy of the hypothesis terms, reporting a missed rate and the first ten misses [13][14]. Because the text handed in is cue text joined with spaces [21], a term that was recognized but landed in the wrong cue still counts as present. So this axis catches dropped and mangled domain terms, not displaced ones. Placement is the timing axis's job, and WER cannot help there at all: it discards every timestamp before it starts [5].
Numbers are collected by a separate regex, `\b\d+(?:\.\d+)?\b`, appended to the found terms [15]. That matches digits. Whether "fifteen" ever becomes "15" depends on what the normalizer emits, and the only normalizer behaviour the source claims is brackets, fillers and contractions [9]. Run your own dose strings through it before you trust the entity count on a medical or financial corpus.
The readability plumbing has one sharp edge worth knowing before you aggregate. `chars_per_second` divides flattened cue length by duration and returns `float("inf")` when the duration is not positive [16], so a single zero-length cue poisons any mean; use a median or a percentile. Line breaks are preserved on load deliberately, because they are what the readability check measures [17]. What the material does not supply is a pass value for either reading rate or median cue offset, so those thresholds are an editorial decision you own and have to defend.
Then pin everything. `jiwer.compute_measures()` is gone in recent versions, replaced by `process_words()` returning a `WordOutput` dataclass, and the author notes most blog posts still use the old name [6] - which is a reliable way to date a tutorial. The instruction is to pin what you install and pin it in CI, because these APIs move between majors [7].
