Processing

The Processing tab governs everything that happens between speech and paste: how fast the live preview updates, when WinSTT decides you've finished a sentence, whether a language model cleans up or rewrites the text, and how the result lands at your cursor.

Realtime preview

While you speak, a fast model streams words into the overlay so you see text immediately; the main model still produces the final, accurate transcription when you stop. These controls tune how that live pass behaves. See Model → Realtime model to pick the model itself.

Smart Endpoint

In Toggle and Wake Word modes WinSTT keeps recording across natural pauses, so it needs to decide when a sentence is actually finished. Smart Endpoint uses a sentence-completion classifier to read the live transcript and judge whether you've completed a thought, then sets the silence threshold accordingly.

Sentence pauses

When Smart Endpoint is off, WinSTT falls back to a punctuation-based heuristic: it looks at how the live transcript ends and waits a different amount of silence before finalizing. These three steppers appear only in Toggle / Wake Word mode with Smart Endpoint off.

LLM cleanup

A language model can reshape your words before they're pasted. There is no master LLM switch — this section holds two features that are configured and toggled separately, both running on a local Ollama model by default so nothing leaves your machine unless you point a feature at the cloud.

Each feature runs only when its own enabled is on and a model is configured for its chosen provider. They pick their provider independently — dictation can stay local on Ollama while transforms call OpenRouter, or vice versa. When both are off, the LLM subsystem is dormant.

Provider

Both features choose one of three providers. Switching provider resets that feature's model selection.

Model and effort

The model field depends on the active provider.

When the selected OpenRouter model advertises support, three optional parameters are sent:

Dictation cleanup

When dictation cleanup is on, every transcription is passed through your chosen presets before it lands at your cursor.

Tone presets (pick one)

A tone sets the overall register of the cleanup. At most one tone preset may be active at a time; choosing a new one replaces the old.

Modifiers (independent)

Modifiers stack on top of the tone in any combination. Some carry an intensity level.

Custom modifiers

Beyond the built-ins you can author your own modifiers, each with an optional intensity hint. They are stored separately and merged into the prompt at the end of the modifier stack. The list scrolls once it passes 7 rows.

How presets compose

The presets you enable are assembled into a single prompt in a fixed order, so the result is predictable no matter what you toggle:

1. Polish base

   A universal foundation (clean punctuation, fix obvious errors) is included once per prompt — the same base every tone builds on.

2. Tone

   Your single chosen tone (formal / friendly / technical / casual) layers its register on top of the base. neutral adds nothing beyond the base.

3. Modifiers

   Each enabled built-in and custom modifier is appended, each with its level or intensity hint where it has one.

4. Translate — always last

   If translate is enabled, it is folded in last with a language-generalization clause, so the whole cleaned result is translated as the final step.

A schema clamp is appended to every composed prompt to keep the model returning only the transformed text.

Warmup banner

For Ollama, the panel surfaces a per-feature banner when the model isn't ready, covering three outcomes:

Playground

Each feature has a playground: type sample text, run it through the composed prompt, and see the result. It has no clipboard side effects and its state is transient — nothing is saved, and closing the tab clears it. Use it to confirm your tone + modifier stack does what you expect before turning the feature loose on real dictation.

Transforms

Transforms reuse all of the provider / model / thinking / tone / modifier machinery above, and add two things: a global hotkey and your own prompt list. They act on the currently selected text rather than dictation.

Each transform entry has these fields:

The Transforms playground works the same as dictation's, with its own per-transform editable prompt and transient state.

Context awareness

Dictation cleanup can read text from the window you're typing into so names and jargon get spelled correctly. These controls live next to the LLM config and disappear entirely when the dictation LLM is not configured — the capture pipeline never runs in that case.

Paste behavior

After the final text is produced (and optionally cleaned), WinSTT pastes it at your cursor. This toggle decides whether it also presses a submit key.

File transcription

When you drop an audio file in for batch transcription, this setting picks one or more output formats. It applies to file jobs only, not live dictation.

Shared LLM infrastructure

These settings are shared by both LLM features — one Ollama instance, one OpenRouter account.

Related