WinSTTText-to-Speech (reference)
The Text-to-Speech settings section — enable, pick a voice and speed, and download the on-demand Kokoro voice pack.
The Text-to-Speech section turns selected text into speech with the local Kokoro-82M ONNX model — 54 voices across 9 languages, all on-device.
This page is the quick settings reference. For the full walkthrough — voice catalog, previews, playback flow, and troubleshooting — see the Text-to-Speech guide.
Controls
falsetts.enabledMaster on/off gate. Turning it on probes whether the voice pack is already cached; if not, you're prompted to download it (see below). The toggle is disabled while a download is in flight.
af_hearttts.voiceA searchable list of 54 voices grouped by language, each with a per-row preview button and
a region badge (US, UK, JP, ZH). The languages in this list — US/UK English, Japanese,
Mandarin, Spanish, French, Hindi, Italian, Brazilian Portuguese — are the only ones local
TTS can speak. Picking a voice auto-updates the language. Default is Heart (US female,
af_heart). If a saved voice ID is missing from the catalog, the first available voice is
selected.
1.0tts.speedSynthesis playback speed, 0.5x to 2.0x in 0.1 steps, shown as 1.0x. Applied by the
Kokoro session itself (not post-processing), so pitch stays natural. A reset button
appears when the value is non-default.
9 languages, local only
The voice picker is the full list of languages local TTS supports — Kokoro is the only bundled engine and it doesn't translate. ElevenLabs (already used for cloud STT) covers far more languages, but cloud text-to-speech isn't available in WinSTT yet.
On-demand voice pack
The Kokoro assets ship separately from the installer and download once on first enable, then cache indefinitely. Total is ~190 MB, split across three parts:
| Part | Size |
|---|---|
| Engine pack | ~50 MB |
| Kokoro model | ~100 MB |
| Voicepacks | ~40 MB |
| Total | ~190 MB |
An install banner under the controls shows phase and percent while the pack downloads, with three controls:
| Control | What it does |
|---|---|
| Pause | Halt the download in place — the partial file is kept so you can resume later. |
| Resume | Continue a paused download from where it stopped. |
| Cancel | Abort and discard the in-flight download. |
Downloaded once
Audio is synthesized fully on-device — nothing is sent to the cloud. After the first download the pack is reused on every launch, so subsequent enables are instant.
Device follows the Model tab
Text-to-Speech runs on the same compute device as your speech-to-text model — there's
no separate TTS device picker. Set it once with Device in the
Model tab (model.device); the spawn layer mirrors that choice
onto the server's --tts-device flag, so Auto runs Kokoro on the GPU (DirectML) where
available and falls back to CPU otherwise.
Related
Text-to-Speech guide
The full walkthrough: voice catalog, previews, playback flow, and troubleshooting.
Model & device
Pick the STT model and the compute device that Text-to-Speech also uses.
Hotkeys
Bind the speak-selection combo and the rest of WinSTT's shortcuts.
LLM cleanup
Reshape dictation with tone presets and custom modifiers.
Hotkeys
The four global hotkeys — push-to-talk, re-paste, read-aloud, and transform — plus the while-held combos that cycle modes and cancel a pass.
Integrations
Add API keys to unlock optional cloud speech-to-text (ElevenLabs) and LLM cleanup (Ollama, OpenRouter). Keys are verified, stored encrypted on your device, and only ever sent to the provider you choose.