Updated October 2026 · Speech recognition and on-device transcription
Multilingual transcription lets creators caption and translate content for wider audiences. Tools vary in the languages they support and in how they handle mixed-language speech. Test each language on your own material, check names and terms, and have a fluent reviewer check translations before publishing.
A creator who works in more than one language can reach far more listeners, but the tools that help are uneven. A transcription model may be very good in one language and weaker in another. Mixed-language speech, where a speaker moves between languages in the same sentence, is harder still. The result is that multilingual work needs more testing and more checking than a monolingual project.
This post explains how multilingual transcription works in general, what errors to expect and how to produce captions and translations that serve the audience well.
Many recognition tools support a list of languages, but the quality for each one varies. Some languages have large amounts of training data and perform strongly, while others have less and perform less well. A tool’s list of supported languages does not tell you how accurate it is for your speaker, your accent or your subject. Test before you rely on it.
Translation is a separate step in most workflows. A transcription tool may offer a translation, but the translation quality varies, and it can miss cultural meaning, idioms and humour. Translation needs its own check.
When a speaker switches between languages, many tools struggle. They may force every word into one language, or they may misidentify a language and produce nonsense. If your content mixes languages, check whether the tool supports multiple languages in the same recording, and test a sample that contains the switches you expect.
Consider marking the switches in the transcript, so a reader can see where the language changed. That helps the editing and the translation.
Names, places and specialist terms are the most common errors in any language. A name may be spelled in several ways, and a tool may choose the wrong one. Build a glossary before you transcribe, with the correct spellings of names, products and technical words in each language. Check the transcript against the glossary during editing.
When you translate captions, keep the timing in mind. A translated line may be longer than the original, and a longer line may not fit the time on screen. Shorten the translation where needed, while keeping the meaning. Use the same caption rules as the original, so readability is preserved in each language.
Keep the speaker’s tone. A formal speaker should not sound casual in translation, and a joke should still be a joke. Those are judgements a fluent reviewer makes better than a tool.
A fluent reviewer should check every translation before publishing, especially for a public video. Machine translation can be useful for a first draft, but it can make errors that change the meaning, and some errors are embarrassing. A reviewer who knows the subject will catch more than a general speaker.
Pay for the review if you can. A small fee for a careful check is often worth more than a faster release with errors.
Keep each language version in a separate file, with a clear name and date. Record which tool produced the first draft and who reviewed the translation. A record makes updates easier, and it helps when a correction is needed later.
The language of the captions should follow the language your audience reads most comfortably, which is not always the language the speaker uses. A creator who speaks a regional dialect may choose a standard form for captions, so more viewers can follow. Decide this deliberately, and say in the description which languages are available.
Some audiences prefer subtitles in their own language even when they understand the original. Ask a few viewers, and let their answers guide the choice rather than assumptions about what they want.
Review takes time, and it is easy to underestimate. A translated video of ten minutes may need an hour or more of careful checking, depending on the content and the reviewer. Build that time into the schedule and the budget, so the review is not the step that gets cut when a deadline approaches.
Keep a note of the reviewer, the date and the version checked. A record of who approved each language is useful if an error is found later.
Language support and quality change as tools are updated, and vendor claims about accuracy are not independent measures. This post describes the general problems, not the performance of any particular product. Test your own languages and your own content before you commit.
Clean audio helps every language, and it is the first step in any multilingual workflow. The Vocal Checker shows clipping, noise and level problems in a recording you own. The Tunedzy Go page covers cleaning a phone recording. For the readability of the captions themselves, read the caption readability post.
Record voice-overs with live noise removal, or clean an existing file. Runs in your browser, nothing is uploaded.
Not necessarily. Test the language on your own material, with your speaker and subject.
Use it for a first draft only, and have a fluent reviewer check the result.
Test a sample with the switches you expect, and mark the switches in the transcript.
Build a glossary of names, places and terms before you transcribe, and check against it.
