---
title: VideoVoice Dramabox
emoji: 🎭
colorFrom: red
colorTo: indigo
sdk: gradio
sdk_version: 5.7.1
app_file: app.py
python_version: "3.10"
pinned: true
short_description: Resemble Dramabox — directable speech for VideoVoice
---

<!--
  ZeroGPU is enabled from the Space Settings UI (not via frontmatter).
  This Space serves Resemble's Dramabox "directable speech engine" via
  POST /api/tools/dramabox. The dub pipeline is reachable but rejects
  voice_mode != "dramabox" (server.py), and the frontend never routes
  dub requests here.

  IMPORTANT — sdk_version is pinned to 5.7.1 to match the upstream
  ResembleAI/Dramabox Space. Reasons:
    - gradio 6.x bundles pydantic >= 2.11
    - pydantic 2.11+ emits bool-shorthand `additionalProperties: True`
      that crashes gradio_client schema parsing
    - Dramabox needs pydantic 2.10.6 (per upstream requirements.txt)
    - That pydantic is incompatible with gradio 6.x
  Bumping this to match the other Spaces (6.12.0) breaks the build.
-->

# VideoVoice — Dramabox

Resemble AI's directable speech engine, mounted as a VideoVoice tool tab.

**Endpoint:** `POST /api/tools/dramabox`
**Frontend:** [/app/dramabox](https://videovoice.app/app/dramabox)

## What's different from the other Spaces

This Space is a **tools-only** Space:
- The `/api/tools/dramabox` endpoint runs Resemble Dramabox against a scene prompt
  (quoted dialogue + stage directions outside quotes).
- Other pipeline endpoints (dub, voice-clone, subtitles, audio-cleanup) are
  defensively reachable but the frontend never routes traffic here for them.

## Prompt grammar

```
<speaker description>, "<dialogue>" <action> "<more dialogue>"
```

- Inside quotes is **spoken**: `"Hello, how are you?"`, phonetics like `"Hahaha"`, `"Mmmmm"`.
- Outside quotes is a **stage direction**: `She sighs deeply.`, `He clears his throat.`
- Avoid writing onomatopoeia (`Sigh`, `Ahem`, `Gasp`) inside quotes — the model will
  speak them literally.

## Setup notes

Required Space Secrets:
- `TTS_ENGINE=dramabox`
- `HF_TOKEN` (same as the other VideoVoice Spaces — for model downloads)
- `LTX_DTYPE=bf16` (optional, matches upstream default)

Required vendored source (committed to the BE repo, deployed via [deploy.sh](https://github.com/Video-Voice/VideoVoice-be/blob/main/deploy.sh)):
- `dramabox_src/` — copy of [ResembleAI/Dramabox `src/`](https://huggingface.co/spaces/ResembleAI/Dramabox/tree/main/src). The `tools_api/dramabox.py` worker adds this to `sys.path` lazily on first request.

## Acknowledgements

Built on [Resemble AI's Dramabox](https://huggingface.co/spaces/ResembleAI/Dramabox). All generated audio is invisibly watermarked with Resemble PerTh.