[Whisper] Conflict of logic in accuracy_eval.py

[wu6u3tw]

Conflict of logic in accuracy_eval.py
2 steps are performed in the script (in order) for the text output of whisper.

1. Filter: only certain characters are accepted from the text output
   e.g. only ‘a-z’, ‘ ’, ‘.’ is accepted

2. Normalize: convert certain phrases into another form
   e.g. “one hundred dollar” -> “100$”.

Issues: Normalize will violate the filtering logic, in the example above
“100$” should be filtered out, but added because normalization comes 2nd

Fix

• Option 1: accept digits and dollar sign: PR filed [Draft] [Whisper] Add labels' in the whisper output #2252

• Option 2: do normalization before filtering
  Risk: might change ref accuracy.

[nvzhihanj]

@mrmhodak @hanyunfan FYI for post-mortem 5.1
@keithachorn-intel

[mrmhodak]

WG meeting: Resolve in 6.0, leave open for now