Choosing a speech mode
Pick a speech engine per dictation mode, not once for the whole app. Local keeps audio on your device; cloud (your own key) or Kalam Cloud trade privacy for speed and accuracy. Match the engine to the job — privacy, offline, or hands-off Max — then set it under Dictation → a mode → Hear your voice.
Decide with this table
| Your priority | Choose | Plan |
|---|---|---|
| Maximum privacy | Local (SenseVoice, Whisper Base, Parakeet, or Moonshine) | Free and up |
| Speed and accuracy | Cloud (BYOK) — Groq or OpenAI | Free and up (needs your key) |
| No API keys to manage | Kalam Cloud | Max only |
| Works offline | Local | Free and up |
| Sensitive apps (password managers, etc.) | Forced Local when detection is on | Free and up — details |
Modes left on Same as Base inherit the Base mode's engine. Background concepts: Speech-to-text.
Recommended setups
- Everyday privacy — Set Base to Local, download an engine in Local models, leave Improve off for plain typing.
- Fast drafting — Set Base (or Email) to Cloud, add a Groq or OpenAI key via BYOK providers.
- Zero key setup — On Max, pick Kalam Cloud for the modes you use most; quotas live under Account → Plan.
- Mixed work — Cloud for Email/Document; Local for Notes or any mode you use in sensitive apps.
If sensitive-app detection matches the focused app, Kalam forces local transcription for that dictation — even if the mode says Cloud. Without a local engine installed, dictation fails closed rather than sending audio to the cloud.
Steps to change a mode's engine
- Open Dictation and click the mode card (or edit Base).
- Under Hear your voice, choose Local, a cloud provider, or Kalam Cloud.
- For Local, confirm an engine is downloaded under Settings → AI & Models → Local engines.
- For Cloud, confirm a key exists under Settings → AI & Models → Providers.
- Dictate once and check the status bar latency and mode name.
Earlier releases offered global Auto and Hybrid speech modes. Prefer per-mode providers now; the status-bar STT picker is a legacy control.