Why 99% Accuracy Still Isn't Enough to Trust Auto-Generated Subtitles

A developer building a subtitle editor discovered that a single misrecognised word — 'HP' transcribed as 'gp' — exposed a critical flaw in how auto-subtitle accuracy is communicated to users. While speech-to-text systems are typically measured using Word Error Rate (WER), that metric counts errors without weighing their real-world consequences, such as a dropped 'not' reversing a sentence's meaning. The developer found that ASR models do return word-level confidence scores, but polished text interfaces often obscure this uncertainty, making every word appear equally trustworthy. To address this, they normalised each transcribed word into a structured editor object capturing text, timing, and a confidence score, with absent confidence values defaulted to zero rather than assumed correct. The core conclusion is that auto-generated subtitles should be treated as drafts requiring human review, and accuracy scores describe model performance — not a finished, reliable product.