WinSTT logoWinSTT
Settings

Model

Choose the speech-to-text model, quantization, and device — plus the realtime preview model and one-pass translation to English.

The Model tab is where you pick the engine that turns your speech into text. It runs two models: a large one for the final, accurate pass, and a smaller one for the live preview you see while you speak.

73
Models across 11 families
5
Quantization formats
2
Models (final + preview)
Local
Default source

Main model

The main model runs the final, accurate transcription pass — the text that actually gets pasted. Everything in this section configures it.

Search the catalog, pick a quantization, and let downloads stream in the background.
The Main Model section: a Local/Cloud source toggle, the model selector, device, model-unload timeout, and a Translate to English switch.
Source, model, device, unload timeout, and Translate to English.
Sourcedefault Local
model.model

Run transcription locally (on-device ONNX models) or via a cloud provider (ElevenLabs, OpenRouter). Cloud is locked until you add an API key in Integrations. Removing a cloud key silently reverts to your last local model.

Modeldefault tiny
model.model

A searchable, filterable list of 70+ models grouped by family. Each row shows its accuracy/speed bars, size, language scope, and per-quantization download badges.

Languagedefault en
model.language

Pin a language or leave it on auto-detect. Hidden entirely for single-language models.

Devicedefault auto
model.device

Auto uses the GPU when available and falls back to CPU; CPU forces CPU-only. Only CPU is offered on machines without a supported GPU.

Model Unload Timeoutdefault min5
global.modelUnloadTimeout

How long the server keeps the model resident after the last transcription before freeing RAM/VRAM. Options run from Immediately to Never.

Translate Todefault ""
model.translateTargetLanguage

Stores the selected target-language code; an empty value disables translation. Supported multilingual models transcribe and translate in one pass without an LLM round-trip.

The model picker

Opening the selector reveals the full catalog. Models are grouped by maker, each row showing its hardware fit, size, the languages it covers, and which quantizations you can download.

The open model picker dropdown: a maker rail on the left and model rows showing accuracy and speed bars, size, language scope, and quantization badges.
Search, filter by family or language, and download a specific quantization per model.

A Suggested filter chip sits next to the search box and is on by default. It hides models that wouldn't fit in your machine's free memory alongside everything else you have enabled (your other STT slot, TTS, and the cleanup LLM), and greys out individual quantizations that would need more RAM or VRAM than is available. Toggle the chip off to browse the full catalog regardless of fit. The same filter appears in the TTS and LLM pickers.

Each model exposes one or more quantizations — compressed weight formats that trade a little accuracy for big wins in size and speed.

The picker only shows quantizations the upstream repo actually publishes; the server falls back to fp32 if a chosen variant is missing.
QuantizationPrecisionRelative sizeBest for
Autofp32 (default)100%Maximum accuracy
fp1616-bit float~50%GPU (DirectML) — fastest
int8 / uint88-bit integer~25%CPU — good balance
q4 / bnb44-bit~13%Smallest download, lowest RAM
q4f16mixed 4-/16-bit~20%Small with better quality than q4

Download, pause, resume

Downloads stream in the background without interrupting the running model, so the app stays connected. You can pause, resume, or cancel any download from its badge, and delete a single quantization to reclaim disk without removing the others.

Realtime model

The realtime model powers the live preview that appears while you speak. It runs on a smaller, faster model so words show up instantly; the main model still produces the final text.

The Realtime Model section with a model selector and an update-interval stepper.
A smaller model and an update-interval stepper drive the live preview.
Realtime Modeldefault tiny
model.realtimeModel

Any model under ~700M parameters. If your main model is already small enough to run live, this slot is locked to it (one model serves both) and the picker is disabled.

Update Intervaldefault 0.02
quality.realtimeProcessingPause

Minimum seconds between live-preview updates. Lower feels snappier but costs more compute.

Requires a restart

Most model changes hot-swap live. Switching the backend family or toggling cloud vs local reconnects the engine, which takes a moment — the overlay shows a brief "swapping" state.

Browse the catalog

All 70+ models across 11 families, with accuracy, speed, and language coverage.

Compute & quantization

What CPU, DirectML, and OpenVINO mean — and how to read the quantization labels.

LLM cleanup

Reshape the transcribed text with tone presets and custom modifiers.

Cloud STT

Use ElevenLabs or OpenRouter instead of a local model.

Settings

The frameless WinSTT settings window for model, audio, processing, hotkeys, LLM cleanup, text-to-speech, and integrations. Changes save immediately.

Audio

Pick your input and output devices, tune the dual-VAD pipeline that decides when you're speaking, and set how the microphone is held between recordings.

On this page

Main modelThe model pickerRealtime modelRelated