Captions without a GPU using Windows Live Captions

Captions can come from two places: from our own caption server (Whisper) or from the Windows Live Captions service, the system's built-in speech recognition. This guide is about the latter. The basics are in the First start guide.

When to use which source

The two sources can be switched while running on the caption window's Incoming audio section settings panel (it opens with the gear icon) or on the Audio source tab of the settings window. If there is a machine with a graphics card on the network, the One GPU machine on the network guide describes that path.

The Live Captions tab

The Live Captions tab of the settings window shows the service's current state on this machine in three levels: whether it is installed, configured and running – with the caption language set and the recognition languages installed. The Settings button opens the Windows setting (Accessibility → Captions), where the language pack can be downloaded; the re-check button asks for a fresh answer.

If it is proven not installed, the source cannot be chosen – neither on this tab nor on the caption window's panel; on the Audio source tab it appears as a disabled item.

Switching it on step by step

1. Open Windows Settings (Accessibility → Captions), set the recognition language and download the language pack.
2. On the Translation tab of the settings window (Languages) give the speech language exactly. A candidate list (hu,en) and automatic detection are not accepted by this source, because the system does not tell what it recognised; switching the source moves to the first candidate by itself. The two settings must match.
3. Switch the source to Windows Live Captions on the Incoming audio section's panel or on the Audio source tab (from the command line the --source windows switch does the same).

The program starts the system's captioner (if it is not running, it asks first), hides its window – recognition runs anyway – and restores it on exit. It leaves a running instance alone: if it was already open, it attaches to it.

How text becomes a sentence

The system gives no "final" signal, so the sentence boundary comes from punctuation. A sentence becomes final (and goes to translation) when newer text already follows it, or the caption has not changed for 1.2 seconds – the system corrects the text afterwards too (capital letters, an inserted full stop). The preview text shows immediately meanwhile.

What you lose compared to the own server

• The model is not ours: no model choice, the recognition quality is that of the Windows language pack.
• The segmentation is not ours: the segmentation controls (mode, profile, timing values) are disabled, because we do not schedule the recognition.
• The language must be given exactly, in advance: no candidate list and no automatic language detection.

Translation and interpreter without a GPU

Translation still needs the translator (translatorserver), independently of the Windows recognition. The translation model needs video memory (16.7 GB measured), so it typically runs on another machine with a graphics card; the caption program reaches it at the address set on the Servers tab. Whether the services are reachable is told by the --check switch:

livecaption --check

The interpreter's voice, however, can speak without a graphics card: the in-process voice (Local mode, with a Piper model) speaks with a fixed voice inside the caption program's own process. It needs neither the voice service nor a graphics card, only a model downloaded once (26–115 MB).

You choose the voices on the Translation tab of the settings window; the list can be filtered by language and quality. The cloned voice of the voice service (XTTS-v2) needs a graphics card, and the model is under a non-commercial licence – and cloning needs the conversation partner's voice, so use it only with their knowledge. The interpreter needs headphones: from a loudspeaker the interpreter's voice would get back into the captioning.

If something does not work, the Troubleshooting guide helps by symptom.