Skip to main content
Kalam Documentation

Speech engines: On-device vs Cloud

You choose where speech recognition runs—on your computer or in the cloud, whichever you like. Kalam treats both as equal options and sets no forced default. Set your shared baseline engine under Dictation → Models, and override individual profiles whenever a job calls for something different.

Speech-to-text selection on Dictation → Models showing Cloud and On this device options. (Click to zoom)

On-device vs Cloud comparison​

ConsiderationOn-device modelsCloud providers (your own API keys)
Where audio is processedEntirely on your computerSent to your chosen cloud provider
Works offlineYes—after downloading the model onceNo—requires an active internet connection
Setup neededDownload a model (103–460 MB) under AI & Models → Local providersPaste an API key under AI & Models → Cloud providers
Password managers & Sensitive appsRuns seamlessly without interruptionAutomatically switched to an on-device model (fails closed if none installed)
PlansIncluded on Free and ProIncluded on Free and Pro
Install at least one on-device model even if you prefer Cloud

If you dictate inside a password manager or private window while Sensitive apps protection is active, Kalam blocks cloud audio automatically. Having an on-device model downloaded prevents failed takes in sensitive windows.

Available engines​

On-device engines​

Download and run models locally through Settings → AI & Models → Local providers:

RuntimeModelDownload sizeLanguagesGuide
Sherpa-ONNXSenseVoice Small150 MBEnglish, Chinese, Japanese, Korean, CantoneseLocal models
Whisper.cppWhisper Base142 MB99 languagesLocal models
Sherpa-ONNXParakeet TDT 0.6B460 MBEnglishLocal models
Sherpa-ONNXMoonshine Small103 MBEnglishLocal models

Cloud speech providers​

Connect your own API keys in Settings → AI & Models → Cloud providers:

  • Groq (speech + AI polish)
  • OpenAI (speech + AI polish)
  • Google AI Studio (speech + AI polish)
  • Deepgram (speech)
  • AssemblyAI (speech)
  • ElevenLabs (speech)
  • xAI (speech)

Max (hosted service)​

Max is listed in the app and on kalamvoice.com/pricing, and is not for sale yet. When available, Max will provide managed speech and AI so there are no local downloads or external keys to manage.

Assign an engine​

  1. Connect a Cloud provider key or download a Local model under Settings → AI & Models.
  2. Open Dictation → Models to choose the shared speech engine used by Plain dictation and inheriting profiles.
  3. To give a specific profile (such as Coding or Notes) its own engine, open that profile in Dictation and change Hear your voice.