How to Transcribe a Video in Spanish, French, or 90+ Other Languages
Step-by-step guide to getting accurate transcripts of foreign-language videos using AI — including handling mixed-language content, dialects, and translation.
Most "video transcript" tools assume your video is in English. If you've ever tried to transcribe a Spanish-language YouTube interview, a Korean K-drama clip, or a Japanese tutorial, you've probably hit one of two walls: (a) the tool returns an English phonetic mess, or (b) it returns nothing at all.
The shift in the last 18 months is that AI transcription models are now genuinely multilingual — not "English plus a few extras", but trained from the ground up on dozens of languages. Here's how to use them well.
Which Languages Actually Work
SubGrab uses OpenAI's Whisper model under the hood. Whisper was trained on 680,000 hours of audio across 99 languages. Accuracy varies by language depending on how much training data was available.
For seven languages besides English there are real measurements. Hugging Face's Open ASR Leaderboard tests whisper-large-v3-turbo — the model SubGrab runs — on each (read on 2 November 2026; words right is 100% minus the word error rate):
| Language | Word error rate | Words right |
|---|---|---|
| Spanish | 3.9% | 96.2% |
| Portuguese | 4.6% | 95.4% |
| Italian | 5.4% | 94.6% |
| Dutch | 5.7% | 94.3% |
| German | 6.1% | 93.9% |
| French | 6.7% | 93.3% |
| Hindi | 41.6% | 58.5% |
The big European languages come close to English, which is 95% on clear speech. Hindi is the outlier: on the leaderboard's Hindi test about four words in ten come out wrong, so expect to edit a Hindi transcript heavily — and if the video already has Hindi captions, SubGrab extracts those instead, which is exact.
For the other 80-plus languages Whisper handles there is no comparable public measurement of this model, so treat any single accuracy figure for them with suspicion — including tiers that look precise. As a rule, languages with plenty of training data do better than those with little, and every language suffers more from noise and accents than English does. The full list Whisper was trained on is in the Whisper paper; try one video before you transcribe a batch.
How to Get the Best Results
Three things matter, in order:
1. Audio Quality > Everything
A clean recording in Tier 3 Bengali will beat a noisy recording in Tier 1 English. Background music, multiple overlapping speakers, low bitrate, or compression artifacts all hurt accuracy more than the choice of language model does.
If the source video has clear single-speaker speech, expect close to the high end of its tier. If it's a chaotic group call with music in the background, expect the low end.
2. Let the AI Auto-Detect Language
You don't need to tell SubGrab what language a video is in. Left to itself, Whisper judges the language from the first 30 seconds, which goes wrong when a video opens with music, a countdown or silence: it guesses English and writes the real speech out as an English translation. So since 13 October 2026 SubGrab doesn't let the opening decide. It samples four stretches spread across the video, lets them vote, and transcribes the whole video in the language they agree on. If they don't agree, nothing is forced. More on this in why an AI transcript can come back in English.
The exception is mixed-language content — videos that switch between two languages mid-sentence. Whisper picks the dominant one. If your video is 70% English with Spanish quotes interleaved, you'll get an English transcript with the Spanish bits transliterated. There's no perfect fix for this; the workaround is to manually transcribe the second-language segments after.
3. Long Videos Are Fine
For recordings over roughly two hours, SubGrab splits the audio into 45-minute parts and transcribes them in parallel. Every part is transcribed in the same language, the one the whole recording voted for, so a long video can't drift into another language partway through or inherit a wrong guess from its opening seconds.
A Worked Example: Spanish YouTube Interview
Let's say you want to transcribe a 25-minute Spanish-language interview from a Mexican YouTube channel.
1. Copy the YouTube URL.
2. Paste it into SubGrab.
3. If YouTube has Spanish captions for the video, SubGrab will pull them in seconds — no AI needed, and 1 credit however long the video. Captions on Spanish-language YouTube are often manually-added by the channel and will be very accurate.
4. If captions aren't available, click "Use AI Transcription" (1 credit). The transcript comes back in Spanish, with timestamps.
5. Download as TXT for reading, SRT for subtitles, or VTT for web playback.
6. (Optional) Run the Spanish text through DeepL or Google Translate for a side-by-side English version.
The whole flow takes about 90 seconds for a 25-minute video.
Special Cases
Chinese (Mandarin vs Cantonese)
Whisper's "Chinese" mode is trained primarily on Mandarin. It will attempt Cantonese videos and the result is usually understandable but imperfect — expect 75–85% accuracy. For Cantonese-heavy content (Hong Kong news, Cantonese vlogs), the transcript will need editing.
Indian Languages
Hindi, Tamil, Telugu, Marathi, Bengali, Punjabi, Gujarati are all supported. Hindi is the most accurate; the South Indian languages tend to have lower accuracy because there's less training data. Code-switched content (English mixed with Hinglish, for example) is a known weak spot — Whisper will pick one language and transliterate the other.
Japanese and Korean
Both work very well, including for casual conversational speech. The transcripts come back in native script (Japanese: kanji + hiragana + katakana; Korean: Hangul). If you need romanised output, run the result through a separate romaniser tool.
Arabic
Modern Standard Arabic works well. Regional dialects (Egyptian, Levantine, Gulf) work but with lower accuracy and the model will sometimes "correct" dialect words to MSA equivalents. For dialect-heavy content, expect to edit.
What About Translation?
SubGrab transcribes in the original language of the video. We don't translate by default, because:
1. Translation quality varies hugely by language pair, and forcing it would silently degrade results
2. Many users want the original-language transcript (for citation, language learning, or to translate themselves with a tool they trust)
3. Modern translation tools (DeepL, Google Translate, ChatGPT) are very good and you probably already have a preference
If you want a translation, copy the SubGrab transcript and paste it into your translator of choice. For long transcripts, DeepL has the cleanest results in our testing for European languages; ChatGPT or Claude works better for Asian languages because it can handle context.
Pricing
Multilingual transcription is the same price as English: 1 credit per video up to 60 minutes (longer videos use 1 credit per 60-minute chunk). Every new SubGrab account gets 2 free credits.
Try it free — paste any video in any language →.
If a YouTube transcript comes back in a language the speaker never used, the video has probably been auto-dubbed — see YouTube transcripts in the wrong language.