Accuracy and languages

INDXR produces text two ways — by reading a video's existing captions, or by transcribing the audio — and how accurate the result is depends on which you pick. At AssemblyAI — the speech-recognition service INDXR uses for AI transcription — accuracy is reported per language, so accuracy and language coverage are the same subject.

Auto-captions

YouTube auto-captions are generated either by the video creator or by YouTube's own speech recognition. INDXR fetches these verbatim — their accuracy depends entirely on the source caption quality, not on INDXR processing. Automatic captions get words wrong (names, brands, technical terms) and get more wrong when the audio is poor, when speakers overlap, when the accent is strong, or when the language isn't English. They are also written for on-screen reading: short lines, no punctuation, no sentence structure.

AI transcription

AI transcription reads the audio itself and writes the transcript from scratch, with punctuation, capitalisation and real sentences. It runs on AssemblyAI Universal-3.5 Pro, and it handles names, jargon and accented speech far better than automatic captions do. It is still speech recognition, though, so it is not flawless: poor audio, people talking over each other, and less widely spoken languages all lower the result. How accurate it is in practice depends mostly on the language — see the bands below.

We don't publish a single headline accuracy figure for INDXR, because a number measured on clean English audio wouldn't tell you what to expect on your video. The per-language bands from AssemblyAI below are the honest guide.

Which languages, and how accurate

For AI transcription, INDXR automatically uses the best model for your video's language — it is a language router, not a fault fallback. AssemblyAI Universal-3.5 Pro natively covers 18 languages; a broader model, Universal-2, covers 99, and the language is detected automatically. Caption extraction works for any language YouTube provides captions for.

AssemblyAI reports accuracy per language as an expected word error rate (WER) — the share of words transcribed incorrectly, so lower is better — grouped into four bands:

• ≤ 10% (high) — English, Spanish, French, German, Dutch, Italian, Portuguese, Japanese, Russian, Swedish, Turkish, Ukrainian, Polish, Indonesian, Catalan.
• 10–25% (good) — including Arabic, Mandarin Chinese, Hindi, Korean, Danish, Greek, Hebrew, Vietnamese, Thai, Finnish, Norwegian.
• 25–50% (moderate) — including Persian, Tamil, Marathi, Swahili, Icelandic, Welsh, Kazakh.
• > 50% (limited) — including Bengali, Gujarati, Telugu, Kannada, Malayalam, Nepali, Burmese.

The lists show representative languages; AssemblyAI's page has the complete breakdown, and these bands can change as their models improve.