How Liveglot works
Liveglot is built so recognition and translation stay separate. Audio goes to the recognizer you choose; only finalized text is sent to your translation model.
1. Choose what to listen to
Before you press Start, pick system audio (whatever is playing on the Mac), the microphone, or both. System audio is the usual choice for a browser tab, Zoom, YouTube, Twitch, or another training-room player.
2. Recognize speech
The recognizer produces partial results while someone is still talking. Liveglot can show those previews in the subtitle list, but it does not translate every partial fragment.
With local Whisper, audio stays on your Mac and is segmented with voice-activity detection (about eight seconds maximum per final cut). With Soniox, audio is streamed to Soniox using the key and region you configured.
3. Finalize a sentence, then translate
When a segment is judged complete, Liveglot locks the original line and sends that whole sentence—plus recent context—to your OpenAI-compatible translation endpoint. Failed translations keep the original and can be retried; the line is not discarded.
4. Notes and chat
Each session writes a Markdown file (default folder under Documents/Liveglot) with timestamps, original, and translation. Click a line to open the side chat with that line quoted. Chat answers are meant to use this session’s transcript, not a general web search.
Limits to expect
You need Screen & System Audio Recording for system sound. Local models take time to load at the start of a recording. Translation quality and latency depend on the model and network you configured—Liveglot does not host those models.
Setup guide · FAQ · Download