Skip to main content
Kalam Documentation

Downloading local models

On-device speech engines transcribe speech completely offline on your computer. Install a runtime under Settings → AI & Models → Local providers, download a model, and assign it to Plain dictation or an individual profile under Dictation.

On-device models are also required if you rely on Sensitive apps, because protected windows never fall back to cloud speech.

Settings → AI & Models with Local providers expanded showing Sherpa-ONNX models and install controls. (Click to zoom)

Supported local engines & models​

ModelRuntimeApprox. downloadLanguagesBest for
SenseVoice SmallSherpa-ONNX~150 MB5 (ZH, EN, JA, KO, YUE)Fast multilingual capture on CPU across English and East Asian languages
Whisper BaseWhisper.cpp~142 MB99 languagesBroad international language coverage offline
Parakeet TDT 0.6BSherpa-ONNX~460 MBEnglish (EN)Highest English accuracy offline when disk space and RAM allow
Moonshine SmallSherpa-ONNX~103 MBEnglish (EN)Lowest disk and memory footprint for quick startup

Install a runtime first—individual model Download buttons stay disabled until its runtime reads Installed. If a runtime card reports Unavailable on your hardware, use the other runtime or a cloud provider.

Install a runtime and download a model​

  1. Install a runtime

    Open Settings → AI & Models → Local providers, expand Sherpa-ONNX or Whisper.cpp, and click Install (or Install runtime). Wait until the status changes to Installed.

  2. Download a model

    Pick a model row matching your language and disk budget, click Download, and wait until the model finishes unpacking.

  3. Start the engine

    Click Start on the model row if it reads Stopped (or let your first dictation load it on demand). You can also Stop, Restart, or Delete model at any time; Uninstall runtime unlocks once all of its models are removed.

Route dictation to your local model​

  1. Open Dictation → Models to make an on-device model the shared engine for Plain dictation and inheriting profiles—or open a single profile card to override Hear your voice.
  2. Choose On-device and pick the downloaded model.
  3. Click into any text field, hold your dictation hotkey, speak a short sentence, and release to verify.