---
title: Playground
description: Every feature of the SDKs, live in your browser. Pick a model, clone a voice, speak with karaoke highlighting, and dictate.
icon: play
---

<Playground />

Every option the SDKs expose is here:

- **Speak**: model, voice, `exaggeration` (Chatterbox), `buffering`, and token `sampling` (temperature, top-k, top-p,
  min-p, repetition penalty), starting from each model's defaults.
- **Transcribe**: model, device (WebGPU or WASM) and the stop timeout, from the microphone or a file.
- **Voices**: record or upload, with automatic or manual cropping and a maximum clip length.

Everything runs on this device: models download from Hugging Face the first time (then come from the browser cache),
voices you add are stored in this browser's IndexedDB, and no audio or text is sent anywhere.

:::note
Text-to-speech needs **WebGPU**. Speech-to-text runs everywhere with Moonshine; Voxtral Realtime needs WebGPU and a
strong GPU (a 2.85 GB download).
:::

The playground is built with the same public API these docs describe: [`useSpeak`](/speech/packages/react),
[`useTranscription`](/speech/packages/react#live-transcription) and [`useVoices`](/speech/packages/react#a-voice-picker), with shared
engines created once per page. Its source is in
[`apps/docs/lib`](https://github.com/kucukkanat/speech/tree/main/apps/docs/lib).
