Auto Generate Subtitles From Any Video, Privately

Choose the spoken language for better subtitles

Auto-detect listens to the opening of the recording and guesses the language. It is usually right, but a short intro with music or a greeting in another language can send it the wrong way, producing a translation or gibberish. Picking the language yourself removes that risk.

• Set the language explicitly for recordings that open with music, silence or a jingle.
• For clips where people switch languages, pick the dominant one and fix the rest by hand.
• Whisper transcribes in the spoken language; for English captions on a Spanish clip, translate the finished .srt.
• Strong regional accents are handled well by Whisper, but uncommon names still need checking.

If your language is not in the list, auto-detect may still recognize it, since Whisper was trained on around a hundred languages, though accuracy varies widely. Widely spoken languages with lots of training audio, such as Spanish, German and Japanese, come out close to English quality, while smaller languages tend to need a heavier correction pass. Mixed-language recordings, like a Hindi interview with English technical terms, usually keep the loanwords but may spell them phonetically.

What decides accuracy, and how to improve it

Recognition quality depends far more on the recording than on the software. A clear lapel microphone in a quiet room can give near-perfect text, while a phone across a noisy cafe will produce gaps and guesses.

• Background music under speech is the biggest single cause of errors; lower it or use the voice-only take.
• Two people talking over each other merge into one confusing block.
• Very quiet or distant speech is often skipped entirely.
• Jargon, product codes and names are spelled phonetically, so plan a correction pass.

A quick test tells you what to expect: transcribe the first minute, count the errors, and multiply. If a minute contains more than five mistakes, clean the audio or switch engines before processing the whole recording. When accuracy really matters, such as legal, medical or research recordings, the separate AI Studio runs larger Whisper models in the browser, up to Large-v3 Turbo, which needs a capable GPU but is much more accurate.

Fix timing and line breaks in the editor

Generated blocks follow the speech, which is not always how people like to read. Each block in the editor shows a number, a time range in hours, minutes, seconds and milliseconds, and the text. You can edit all three.

• Split a long block at a natural pause and adjust the two time ranges so they do not overlap.
• Extend a block that disappears too fast; about one second per five words reads comfortably.
• Merge tiny fragments like yeah or okay into the neighboring block.
• Keep numbering sequential if you add blocks; most players tolerate gaps, some strict ones do not.

Once the text is clean, download .srt for YouTube and most editors, .vtt for web players, or burn the captions into the picture for social apps.

Speed, long files and the optional API key

On a recent laptop the in-browser model handles a 10 minute clip in a few minutes; older phones take noticeably longer, and very long files can run out of memory. Trim the section you need first when you only want part of a recording.

• Keep other heavy tabs closed while transcribing a long file.
• Phones are limited to roughly 250 MB per file, desktops to about 2 GB.
• In the AI Key tab you can add your own Groq or OpenAI key for much faster, larger-model results.
• With a key, the extracted audio is sent to that provider, so skip it for confidential material.

For foreign-language viewers, drop the finished .srt into the AI Translator, which handles 35 languages and keeps every timestamp. That step uses cloud AI, unlike the transcription itself.

Step-by-step

Common mistakes to avoid

Pro tips

Frequently asked questions