Settings: Speech Model

Engine Fields

Choose a model in Settings → Speech model. Expand Advanced setup and diagnostics for local paths, thread count, VAD, and Test setup.

Fields:

• whisper-cli path
• model path
• thread count (1 to 16)

For DMG installs the paths are auto-populated from the bundled runtime under the app bundle (Steno.app/Contents/Helpers/whisper-cli and Steno.app/Contents/Resources/WhisperModels/...). For source builds, point both fields at your local whisper.cpp build and the model file you downloaded.

If a path field shows File not found at this path, the file is missing or moved. Fix the path or reinstall the DMG. A speech model or transcription tool that is missing at launch is reported in Last activity on the Dictate tab and points to Settings → Speech model.

Model Library

The Model Library under Speech model shows the included model and optional downloads:

• small.en — the bundled English default
• medium.en — an optional larger English model
• large-v3-turbo — an optional multilingual model; compare results on your Mac

Each model has one or more status labels:

• Included — bundled with the DMG (Small only).
• Downloaded — present locally on this Mac.
• Using — currently the active model.
• Recommended — picked by the hardware compatibility matrix.

Use the button beside a model to Use a downloaded model or Download one you don't have yet. Downloads come from the whisper.cpp model repository on Hugging Face. You do not need shell commands.

Each download is checked against the published file before it is installed. A download that doesn't match, such as a page returned by a network filter, is deleted, the current model stays in use, and Steno says what happened. A failed download shows the reason next to the Download button. Downloaded models can be removed or downloaded again; removing the model in use switches to the included Small model first. Downloading a model keeps a voice-detection model you chose yourself. base.en is also recognized through advanced paths but is not an upgrade listed here.

Hardware Recommendation

Speech model reads your Apple silicon chip class and unified memory and shows:

• detected hardware (for example M3 Pro · 32GB unified memory)
• the recommended model from the built-in compatibility matrix
• a status line for the currently active model, marked as one of:
  • Validated — measured against the canonical release-signoff benchmark for this exact row
  • Warning — recommended default for this hardware tier, but not yet release-validated
  • Custom — model file is outside the curated set (custom or quantized)

Macs with an M1, M2, or M3 Ultra chip also get a recommendation. The matrix includes dated test results, such as the April 23, 2026 m5-pro / 64GB / large-v3-turbo result. Validation applies to the exact tested combination. For other combinations, use the recommendation as a starting point and compare results on your recordings.

Save or discard a settings draft before switching models. Model selection uses its own action. Downloads need internet access; installed models can be used offline. The first recording loads the selected model and later recordings reuse it when possible. Changing a transcription setting, such as the model or thread count, reloads the model; saving other settings keeps it loaded.

Test Setup

Test Setup transcribes a short test clip with the current settings, using the main engine and the fallback tool, and reports each step. It doesn't affect a dictation in progress or add anything to History or Insights.

To check recording and insertion, follow Test Setup with a dictation in your editor.

VAD

Voice Activity Detection is enabled by default when a VAD model is available.

• DMG install: Silero is bundled, and VAD is enabled by default.
• Source build: download the VAD model with ./download-vad-model.sh silero-v6.2.0 from vendor/whisper.cpp/models and either point Steno at it directly or place it next to your selected Whisper model so the path is derived automatically.

If the optional VAD model is missing, Steno keeps your other Whisper paths valid and shows a plain instruction, with a button to use the included model.

Performance Guidance

Thread count affects speed and system load.

• Start with the default thread count and only change it after you have a baseline.
• Compare changes on the same recording; more threads do not guarantee lower latency.
• Lower the count if your Mac gets hot or its fans become noisy during dictation.

After any thread count change, dictate a typical-length sample and compare transcription time and battery use.

Common Engine Errors