Captions make your stream accessible to people who are hard of hearing, watch without sound or do not speak your language. Thanks to local speech recognition this now works without a subscription and without your voice leaving your PC. A few decisions determine how well it works in the end.
Two ways into the picture
Captions can be part of the video image or switched on by the viewer. Both have their place:
| Way | Advantage | Drawback |
|---|---|---|
| In the image via OBS (browser or text source) | Everyone sees them, also in clips and recordings, on every device. | Nobody can turn them off; they cover part of the picture. |
| Switchable via Twitch extension (video overlay) | Whoever does not need them does not see them. Several languages are possible. | Only where Twitch shows extensions; not in the recorded image. |
Twitch has three anchors for extensions: a panel below the stream, a video overlay across the whole video and a component in part of the video. Only the overlay suits captions. Our own extension therefore shows captions only in this anchor.
Speech recognition: local or cloud
At the core is a program that turns your voice into text. There are two approaches:
- Local: models such as Whisper run on your PC, on the CPU or the GPU. The audio does not leave your PC. The models have to be downloaded once; depending on size they range from a few hundred megabytes to over a gigabyte.
- Cloud: a service on the internet recognises the speech. That takes load off your PC, but your audio goes to a third-party provider, and it usually costs per minute.
Larger models recognise more accurately but need more computing power. A small model is a good start. Only when it makes too many mistakes is a larger one or a GPU worth it.
How fast it is – measured
Captions always arrive a little after the spoken word. We measured this for our tool with a test recording: Whisper with the “small” model, German speech, CPU only.
- Delay until the finished text: 0.9 to 1.6 seconds.
- Word error rate: 4.1 percent, so roughly every 25th word wrong.
- Translation into three languages with a local translation model: 330 to 390 milliseconds per sentence.
These values come from a test file, not from a real microphone with background noise. The path to the viewer’s screen comes on top. Measure on your own setup before you promise any numbers.
Translation: helpful, but label it
The recognised text can be translated into further languages, also locally with a dedicated translation model or a language model on your own PC, or through a service. Machine translation is good enough to follow a stream, but not flawless. Puns, names and jargon from your game often go wrong.
So label translated captions as machine-translated. Our extension shows an “AI-translated” badge for this. That way everyone knows an odd sentence did not come from you.
Readable, not just present
- Large enough: many people watch on their phone. What looks big on a monitor is tiny there.
- Contrast: light text on a semi-transparent dark background, never directly on colourful gameplay.
- Two lines at most: nobody reads more text before the next line arrives.
- Keep space free: captions do not belong over health bars, maps or chat overlays.
- Local access only: if the tool serves the captions as a small web page for the browser source, it should only be reachable on your own PC. Our tool listens only on
127.0.0.1for this.
Data protection: your guests speak too
Speech recognition captures everything in the audio, including teammates in voice chat and guests. With local recognition the audio stays on your PC; only the finished text is passed on, for example to the extension. With cloud recognition the audio itself goes to a provider. Tell your guests beforehand that a transcript is being made, and name in your privacy notice which service receives the data.
Checklist
- Captions burned in, switchable via extension, or both?
- Local recognition if the audio should not leave your PC?
- Model size suited to your PC, delay measured yourself?
- Translations labelled as machine-translated?
- Large text, high contrast, two lines at most?
- Guests informed and privacy notice updated?