AutoCap transcribes speech with Whisper models that run inside your browser, in 56 languages or with automatic detection. There are two modes, Fast and Accurate, and the right choice depends mostly on your language and on how much the final accuracy matters compared with waiting time. This guide explains what changes between them, so you can decide before you start rather than after a slow run.
The two modes at a glance
| Mode | Model | One-time download | Best for |
|---|---|---|---|
| Fast (default) | A small Whisper model chosen for your language | About 300 MB | Most videos in widely spoken languages; quick on most devices |
| Accurate | Whisper Large v3 Turbo, multilingual | About 760 MB | Lower-resource languages, difficult audio, and final exports where accuracy matters most |
In testing, the large model was far more accurate than the small ones, but on an ordinary laptop graphics chip it took around ten times longer. That gap is the reason Accurate is opt-in rather than the default: for a clear English voice-over, Fast is usually all you need, while for a language the small model handles poorly, the extra wait is the only way to get usable text.
Languages that get a dedicated model
For three languages, the Fast mode does not use the general multilingual model. A model trained for that language was compared with the multilingual small and large models on public FLEURS test clips, and it is used only where it held up.
- English uses an English-only build of Whisper Small. Dropping the other languages makes the model more focused for English speech.
- Swedish uses KB-Whisper Small from KBLab at the National Library of Sweden. On the test clips it made about a third of the errors of the multilingual small model. Its export has no word-level alignment, so words are timed inside each caption segment instead of one by one.
- Norwegian uses NB-Whisper Small from NbAiLab at the National Library of Norway, with word-level timestamps.
Every other language uses the multilingual Whisper Small in Fast mode and the large model in Accurate mode.
Languages where the small model struggles
Whisper was trained on far more audio in some languages than in others. For lower-resource languages the small models make many more mistakes. The studio suggests Accurate mode when you pick any of these:
Afrikaans, Arabic, Bengali, Welsh, Estonian, Persian, Gujarati, Hindi, Icelandic, Kazakh, Kannada, Lithuanian, Latvian, Malayalam, Marathi, Nepali, Punjabi, Swahili, Tamil, Telugu, Thai, and Urdu.
From our own testing
Hindi and Persian came out unusable on the small model. If your video is in one of these languages, use Accurate from the start and expect a longer first run, including the larger download.
Pick the language, or let it detect
Auto-detect asks the model to guess the language from the audio. It works well for long, clear speech in a common language. Choosing the language yourself is more reliable for short clips, for videos with background music over the first seconds, and for languages that sound similar. It also lets the studio apply the right dedicated model. If your video mixes languages, transcribe the main one and fix the rest by hand.
Word-level timing and highlight styles
Most of the models AutoCap uses expose the information needed to give every word its own timestamp. That is what lets word-highlight and karaoke-style caption motions land on the right syllable. When a model only provides timing per segment, as with the Swedish model, AutoCap spreads the words across the segment, which is accurate enough for normal captions but less precise for fast highlight animations.
Getting better results from either mode
- Start with the cleanest audio you have. Background music at a low level costs accuracy; a clean voice track costs none.
- Set the spoken language instead of relying on auto-detect for clips under half a minute.
- Read the transcript against the audio once and fix names, brands, and numbers first, since those are the most common errors in every mode.
- Run Fast first to check the video and the language, then switch to Accurate for the final version if the text needs more cleanup than you want to do by hand.
- On a phone or an older laptop, prefer Fast and keep clips short; the Accurate model is a heavy download and a heavy run.
Which should I use?
- 1Widely spoken language, clear audio: Fast.
- 2Swedish, Norwegian, or English: Fast, which already uses a dedicated model.
- 3A language from the list above: Accurate.
- 4Interviews, strong accents, or technical vocabulary where errors would be embarrassing: Accurate, then proofread.
- 5Not sure: run Fast, read the first minute, and switch if it is clearly wrong.
No automatic transcript is perfect in either mode. The point of the editor is that fixing the remaining errors is fast, with the text next to the video and the timeline updating as you type.