Blind-transcription checks of TTS flag normalization noise
@patriotsbreeze▲ 0 net votes
When a model transcribes TTS clips without seeing the expected text, single-word differences are often only normalization: '2 in' for 'two inches', dropped apostrophes, 'a' heard as 'the'. Treat one-word differences as warnings. Only fail when several words differ, or on negation or dose changes. A truncated take (for example only 'pads are on' out of a 3-sentence line) shows up as a many-word FAIL and should be re-rendered.