YouTube Transcripts in Non-English Languages — What Works

If you're extracting transcripts from non-English YouTube videos, there's something you should know upfront before you spend time on a workflow that won't give you what you expect.

What Caption Extraction Gives You for Non-English Videos

YouTube's auto-caption system generates captions in the video's original language. Arabic videos get Arabic captions. Spanish videos get Spanish captions. That much is straightforward.

The problem is at the infrastructure level. When our system downloads captions via YouTube's timedtext API, YouTube's CDN forces the output through an English translation layer — regardless of what language was requested. The URL parameter tlang=en is appended by YouTube's server, not by us, and it isn't overridable through standard API calls.

The result: you submit an Arabic video, you get an English-translated transcript. The language field in the metadata will correctly say "ar" — that's the audio language — but the text itself is the English translation.

This is a YouTube infrastructure limitation, not something unique to INDXR.AI. We've confirmed the same behavior across other transcript tools including Tactiq and YouTubeToTranscript.

If you need the original language text, caption extraction is not the right route. AI Transcription is.

What AI Transcription Gives You

AI Transcription downloads the video audio and runs it through AssemblyAI's speech recognition models directly — bypassing YouTube's caption system entirely.

For Arabic, Spanish, Portuguese, Turkish, Indonesian, and 95 other languages, AssemblyAI's Universal-2 model transcribes the audio in the original language. For English, Spanish, German, French, Portuguese, and Italian, Universal-3 Pro is used — the higher-accuracy model.

The output is the actual spoken language, correctly transcribed, with punctuation. Here's a real example from an Arabic lecture video (Dr. Tariq Al-Suwaidan, 28.5 minutes):

{
  "extraction_method": "assemblyai",
  "language": "ar",
  "segments": [
    {
      "text": "كثير من الناس يُخصّصون كل جهدهم ووقتهم فقط للبحث عن المال",
      "start_time": 35.2,
      "end_time": 42.1
    }
  ]
}

Correct Arabic text. Correct timestamps. The same structure as any English transcript.

When to Use Each Approach

For RAG pipelines specifically: if you're building a knowledge base in Arabic, Turkish, or Indonesian, AI Transcription is the only reliable route to original-language chunks.

Cost

AI Transcription: 1 credit per minute, minimum 1 credit.

A 28-minute Arabic lecture: 28 credits. At Basic pricing (€6.99/500 credits), that's €0.39.

AssemblyAI Universal-2 (used for non-English languages outside the Universal-3 Pro set) has comparable accuracy to Universal-3 Pro for most languages. For Arabic specifically, it handles Modern Standard Arabic and many dialects reliably.

For the full JSON export schema, see YouTube Transcript JSON Export. For audio file uploads, see Audio Upload. For credit packages, see the pricing page.