Spellcheck can't fix "chai GPT": building an LLM caption checker on $5 a month

Описание и применение

Misheard использует LLM для обнаружения и исправления ошибок в автоматических субтитрах YouTube, повышая точность до 78 % и оставаясь в пределах бюджета $5 в месяц.

Задача

YouTube auto‑captions frequently mishear words, producing real‑word errors that standard spellcheckers cannot detect.

Результат

Misheard improves real‑word error detection from 6/79 to 22/79 on hand‑checked videos, with 0.78 precision, and operates within a $5/month budget.

Рабочий процесс

• Upload SRT/VTT file
• Local detectors flag non‑word and phonetic errors
• Chunk transcript (~400 words) and send to LLM with hints
• LLM returns misheard words and suggested fixes
• User reviews and accepts/rejects suggestions
• Export corrected transcript

Доказательства

Over half were real‑word errors (55), and only 41 were non‑words.

Источник

Read-through: The transcript goes out in chunks of about 400 words. Each chunk carries the local detectors' flags as hints, plus priming terms.

Источник

That’s 3.7x and 8.3x the real‑word catches, with better precision, at about $0.05 per hour of audio.

Источник