flaviusburca/DramaboxTTS

DramaBox TTS

Expressive text-to-speech with voice cloning, built on LTX-2 by Lightricks.

DramaBox is Resemble AI's expressive TTS model trained on top of the LTX-2.3 audio branch.

Voice backends

A Voice backend selector switches between two TTS engines, both kept warm:

• —DramaBox (LTX-2) — expressive scene-prompt TTS with stage directions and paralinguistics (laughs, sighs, gasps).
• —Higgs Audio v3 (4B) — fast zero-shot voice cloning via `bosonai/higgs-audio-v3-tts-4b` (ported to plain 🤗 transformers by `multimodalart/higgs-audio-v3-tts-4b-transformers`).

Usage

DramaBox: write a scene prompt with dialogue in "double quotes" and stage directions outside:

A woman says warmly, "Hey, just saying hi, hope you're doing well!"
She laughs softly, "Hehehe, we should do something soon."

Higgs Audio v3: just type the plain text you want spoken — no special formatting needed.

Either way, optionally pick a voice from the gallery to clone a specific speaker. When you do, its preview clip is auto-transcribed on CPU (Whisper, never touches the GPU) into "Reference transcript" — Higgs Audio v3 uses this to improve cloning quality. Edit or clear it as needed.