Friskription → Blog → Live stream translation

Live stream translation: every method, compared

Start here 17 September 2026

There are four ways to watch a live stream in a language you do not speak. Three of them need something the stream has to give you first. This is what each one does, where each one stops, and how to tell in about ten seconds which one you are stuck with.

Why a live stream is harder than a video

A recorded video can be processed in advance. Something can read the whole file, work out where the sentences end, and produce subtitles that are correct before anyone presses play. None of that is available on a live stream: the sentence you are translating is not finished yet, and the next one has already started.

That creates one trade-off you will meet in every tool on this page. Wait longer before showing a line and it reads better but lands later. Show it sooner and it arrives on time but in fragments that get rewritten in front of you. There is no setting that escapes this; there are only different places to sit on the curve.

The second problem is that most live streams carry no text at all. A recorded YouTube upload usually has automatic captions. A Twitch channel almost never does, because captions there only exist if the streamer runs captioning software themselves, and most do not. This single fact decides which of the four methods below is even available to you.

The four methods

1. Caption translators

Browser extensions and built-in features that take an existing caption track and translate it. YouTube's own auto-translate is the best-known example: pick a language in the subtitle menu and the captions appear in it.

What it needs: a caption track. That is the entire story. If the subtitle menu is empty, a caption translator has nothing to work with and will show you nothing — not an error, usually, just silence.

Where it stops: quality is inherited. If the automatic captions mishear a name, the translation faithfully translates the mistake. And auto-captions are generated for far fewer languages than most people expect, so a stream in a language YouTube does not auto-caption has no track to translate even on YouTube.

2. Captions the streamer provides

Some broadcasters run captioning on their end — a hardware captioner, a speech-to-text service wired into their production software, or a human stenographer for larger events. When this exists it is usually the most accurate text you will get, because it was set up by someone who knew the vocabulary.

What it needs: the streamer to have done it, and to keep doing it. You cannot turn it on from your side.

Where it stops: it is nearly always in the spoken language only. A Japanese stream with Japanese captions is a great deal more accessible than one without, and still unreadable if you do not read Japanese — though it does give a caption translator something to work from.

3. Chrome's Live Caption

Built into Chrome, free, and it needs nothing from the stream: it listens to the audio playing on your machine and transcribes it locally. Everything stays on the device, which is genuinely the best privacy story on this page.

What it needs: the feature switched on in Chrome's accessibility settings, and a source language it supports.

Where it stops: the set of source languages is limited, the caption appears in a small fixed panel rather than on the video, and translation is a general-purpose machine translation rather than something reading the conversation for context. It is an accessibility feature that happens to be useful here, not a translation tool. We have measured the gap in detail in a separate comparison.

4. A transcribing overlay

The same idea as Live Caption — listen to the tab's audio rather than look for text — carried further: transcribe the speech with a production speech-recognition model, translate it with a language model that reads each line against the ones before it, and draw the result over the video as a subtitle you can move and restyle. This is what Friskription does.

What it needs: a tab that plays sound. Nothing from the streamer, nothing from the platform, no caption track.

Where it stops: it runs in Chrome, the transcription happens on a server rather than on your machine, and beyond the free weekly allowance it is metered by the minute. It also cannot help with audio it cannot hear — a stream you are watching on a television is out of reach.

Which one to use

A quick way to decide: open the stream and look in the subtitle menu. If there is something in it, start with a caption translator. If it is empty — and on Twitch it will be — the audio is the only text there is, and only something that transcribes can help.

What to expect from any of them

A delay of one to three seconds

Unavoidable: a line has to be spoken before it can be heard, heard before it can be transcribed, and transcribed before it can be translated. Tools differ in how they spend that budget. Friskription offers three presets — Quick, Balanced and Precise — that trade how complete each line is against how soon it lands, and a Sync mode that holds playback back a few seconds so the subtitle arrives with the words instead of after them.

Names will be the weak point

Every speech recogniser is worst on proper nouns, because a name is exactly the word it has the least context for. Streamer handles, character names, in-jokes and game terms are where you will notice it. A glossary that pins a spelling for the session is the only real fix, and is worth setting up once for a channel you watch often.

Noise costs accuracy

Published benchmarks are measured on clean recordings. A live stream has game audio, music, overlapping speakers and a microphone in a bedroom. Every engine scores worse here than on its benchmark, which is worth knowing before comparing two numbers that were both measured somewhere else.

Setting up the transcription route

If the subtitle menu was empty, this is the route you are on. It takes about a minute.

1. Install the extension and sign in

   Add Friskription to Chrome from the Web Store and sign in with Google. The account exists to meter minutes; it opens with thirty free minutes a week, refilled every Monday, and needs no card and no AI provider accounts.

2. Pick the language to translate into

   Fifteen are offered: English, Japanese, Chinese, Korean, Spanish, French, German, Portuguese, Russian, Arabic, Hindi, Italian, Vietnamese, Thai and Indonesian. The spoken language is detected for you, so there is normally nothing else to set.

3. Open the stream and press start

   Subtitles appear over the video within a few seconds. Drag the overlay somewhere it does not cover the action, resize it, and set the font size once — the choice is remembered per site.

What it costs

Methods one to three are free. Method four is metered because the transcription and translation run on someone's servers and cost real money per minute of speech.

• Thirty minutes every week, free, refilled each Monday whether or not you have ever paid. No subscription of any kind.
• A one-time $1.99 starter pack adds 90 minutes and permanently unlocks the Premium features, including the Sync presets.
• Regular packs run $2.99 to $24.99, up to 1,440 minutes. Bigger packs cost less per minute and minutes never expire.
• Or connect your own OpenAI, Anthropic or ElevenLabs API key and pay only your provider — in that mode audio goes from your browser straight to them and never reaches our servers.

Only speech is billed, so the silence between rounds, the waiting screen and the music while the streamer is away do not count against your balance.

Questions