Whisper and Soniox setup
Recognition and translation are configured separately. Pick one recognition engine per session; translation always uses an OpenAI-compatible endpoint you control.
Local Whisper (on-device)
Install whisper-server (for example via whisper.cpp / Homebrew) and point Liveglot at a ggml/gguf model. The app scans common folders such as MacWhisper and whisper.cpp model directories; you can also browse to a file.
The first start of a recording may spend time loading the model. Status text covers loading, ready, and recognizing. Keep enough free memory for the model size you choose.
- Needs a working whisper-server binary path
- Needs a compatible local model file
- Audio for recognition stays on the Mac when this engine is selected
Soniox (cloud)
In Settings → Recognition choose Soniox, paste your Soniox API key (stored in the macOS Keychain), pick a region (US / EU / Japan / India) and model (Real-time V6 is recommended; V5 is deprecated), then use Test connection.
Audio leaves your Mac toward Soniox for that session. If the region and key do not match, connection tests fail—fix the endpoint rather than retrying blindly.
Translation (OpenAI-compatible)
Under Translation & chat, set Base URL, model name, and API key. Paths are joined as you typed them—include /v1 when your provider requires it. You can keep several profiles and reuse the translation profile for chat or choose another.
“Test connection” validates the fields you entered. Saving applies to later translation requests for the current session’s subsequent sentences.
Troubleshooting
Missing whisper-server or model paths show explicit errors in the UI. Soniox auth or HTTP errors surface as connection failures. Permission errors for Screen Recording usually need a full quit after enabling the toggle—multiple Liveglot copies with the same bundle ID can confuse TCC.
- Apple native recognition is not offered in Settings yet—do not look for it as a shipping option
- Windows setup does not apply; there is no Windows build to configure
FAQ
Do I need both Whisper and Soniox?
No. Choose one recognition engine. Translation still needs its own OpenAI-compatible key unless you only keep originals without translation.
Setup guide · FAQ · Download