Speech-to-text
Kalam turns your voice into a transcript with a speech-to-text engine. You choose the engine per dictation mode — there is no single global mode to get right.
The three kinds of provider
| Provider | Where audio goes | Needs a key? | Plan |
|---|---|---|---|
| Local | Nowhere — runs on your device | No | Free and up |
| Cloud (BYOK) | Groq or OpenAI | Your own API key | Free and up |
| Kalam Cloud | Kalam's hosted Whisper | No — uses your license | Max only |
Set a mode's speech engine in Dictation → a mode → Hear your voice. Modes left on Same as Base inherit the Base mode's engine.
Local
On-device engines: SenseVoice Small and Parakeet / Moonshine (Sherpa-ONNX), plus Whisper Base (whisper.cpp). Audio never leaves your machine. Models download once and work offline. Download and manage them in Settings → AI & Models → Local engines. Parakeet is CC-BY-4.0 (NVIDIA NeMo).
Cloud (bring your own key)
Audio is sent to Groq or OpenAI for fast, high-accuracy transcription. You supply your own API key in Settings → AI & Models → Providers. Keys are stored on your device.
Kalam Cloud
On the Max plan, Kalam transcribes for you on hosted infrastructure. No API key to manage; your license is the credential. Monthly usage caps apply (Account → Plan).
Choosing an engine
| Priority | Pick |
|---|---|
| Maximum privacy | Local |
| Speed and accuracy | Cloud (BYOK) or Kalam Cloud |
| No setup, no keys | Kalam Cloud (Max) |
| Works offline | Local |
See the full decision guide: Choosing a speech mode.
Sensitive apps override everything
If sensitive app detection is on and the focused app matches your list, Kalam forces local transcription for that dictation regardless of the mode's provider — and fails closed (no transcription) if no local engine is installed. Details: Sensitive apps and privacy.
Earlier releases offered global Auto/Hybrid speech modes. These are retired: you now pick a provider per mode, and sensitive-app force-local applies to any mode automatically.