Inky ClipsIf you speak Latvian, Lithuanian, Estonian or another language with a couple of million speakers, you’ve probably noticed that “auto captions” often means “auto captions in English”. Many apps don’t offer your language at all, and the ones that do can be noticeably worse at it. This guide explains why, what to expect from the speech model Inky Clips uses, and the habits that make the biggest difference to the result.
Why smaller languages are harder
Speech recognition models learn from recordings paired with transcripts. Whisper, the model behind Inky Clips’ captions, was trained on hundreds of thousands of hours of audio, but the share for each language roughly follows how much of it exists online. English has an enormous amount; Latvian has a tiny fraction of it. The model still learns the language, and well enough to be genuinely useful, but it has heard fewer voices, fewer accents and fewer rare words, so it makes more mistakes.
In practice, for clear Latvian speech you can expect most words to come out right, with errors concentrated in a few predictable places: names, place names, brand names, English loanwords, numbers and the endings of words that are spoken quickly. It’s a draft that takes minutes to fix, not a finished transcript.
Choosing the language matters
When you click Auto captions, the editor asks for the Spoken language rather than guessing. Type “Lat” in the search box and pick Latvian. The editor remembers it for next time and puts it under Last used. Telling the model the language up front avoids a whole class of errors, where a short or quiet opening makes it guess a neighbouring language and carry on in the wrong one.
What helps most
1. Clean sound beats everything
The single biggest factor is the recording, not the language. A clip-on or phone microphone close to the speaker’s mouth, in a quiet room, will caption better than a camera microphone across the room. Wind, echo and traffic all cost accuracy, and they cost more in a language the model knows less well.
2. Caption before you add music
Background music makes speech harder to pick out. If you edit elsewhere, export a version with the voice only, caption that, and add music afterwards, or keep the music bed low.
3. One language per video
The model transcribes in the language you chose. If a speaker switches into Russian or English for a sentence, that part may come out garbled or oddly translated. For mixed videos, pick the main language and expect to fix the other bits by hand.
4. One speaker at a time
Overlapping voices are hard in any language. Interviews where people take turns caption far better than lively group conversations.
Proofreading quickly
Play the video once from the start with captions on. When you see a mistake, pause, click that caption on the timeline and fix it in the Caption text box. Check these first:
- Names and places. “Ventspils” may come out fine; a small village or a surname often won’t.
- Long and short vowels. The difference between a and ā, i and ī is easy to mishear, and it changes the meaning.
- Numbers and dates. They may be written as words or digits; make them consistent.
- Loanwords and brand names, which can be spelled the Latvian way or the original way.
Word-by-word captions: how to make them and edit them all at once covers caption editing in detail, including changing the style of all captions at once.
Diacritics, written by hand
Many handwriting fonts only contain the basic Latin alphabet, which is why apps often drop or replace letters like ā, č, ē, ģ, ī, ķ, ļ, ņ, š, ū and ž. Inky Clips builds the missing accented letters from the base letter and a pen-drawn mark, and writes the mark after the letter, the way you would by hand. The same goes for Lithuanian ą, ę, ė, į, ų, Estonian õ, ä, ö, ü and Polish ś, ź, ż. Try it here:
The clip shows “Sveiki, Rīga!” written in Cursive.
Ink: White
Your audio stays with you
Because the speech model runs in your browser, your recordings never go to a server, which matters when the video is an interview or a client project. The model downloads once (about 1.1–1.6 GB) and is reused for every language. How to add subtitles to a video without uploading it explains how that works.