Dramabox

Updated
24.08.2026
Audio
Run

Dramabox is Resemble AI's expressive TTS built on the LTX-2.3 audio branch: the prompt controls voice, emotion and delivery, with 10-second voice cloning.

At a glance

  • License: LTX-2 Community License
  • Parameters: 3.3B audio-only Diffusion Transformer, IC-LoRA fine-tune of LTX-2.3
  • Language: English
  • Minimum hardware: about 24 GB VRAM peak with the warm server
  • Extras: voice cloning from a 10 second reference, Perth watermark on outputs

What is Dramabox?

Dramabox is Resemble AI's expressive text-to-speech model, built on the audio branch of Lightricks LTX-2.3 under the LTX-2 Community License. It is an IC-LoRA fine-tune of the LTX-2.3 3.3B audio-only Diffusion Transformer with flow matching, conditioned on Gemma 3 12B text embeddings. The design idea is that the prompt controls everything: speaker identity, emotion, delivery, laughs, sighs, breaths, pauses and transitions. An optional 10-second voice reference clones the target timbre.

SpecificationDramabox
Base modelLightricks LTX-2.3, audio-only branch
Parameters3.3B Diffusion Transformer, flow matching
Text conditioningGemma 3 12B embeddings
Voice cloningOptional 10+ second reference, 3 to 30 second window
LanguageEnglish
Release dateApril 2026
LicenseLTX-2 Community License

How Dramabox prompting works

The prompt format splits speech from direction. Anything inside double quotes is spoken literally, including phonetic vocalisations like Hahaha and Hmm. Anything outside quotes is a stage direction the model performs but never says: she sighs deeply, a long pause, his voice cracks. The card's sample outputs run from a regal queen moving from cold fury to a whisper, to a football commentator narrating a trip to the fridge, and the same mechanics drive them all.

Generation is controlled by a handful of parameters: cfg_scale for how strictly the output follows the prompt, stg_scale for expressive emphasis, and a duration multiplier or an explicit target duration for long scenes. Every output carries the Resemble Perth neural watermark by default; it survives MP3 and AAC compression with close to 100% detection accuracy and can be disabled with a flag.

Dramabox hardware requirements

Resemble AI lists about 24 GB of VRAM peak with the warm server, and about 2.5 seconds per generation on an H100 once warm. The download breaks into three parts:

FileSizeContents
dramabox-dit-v1.safetensors6.6 GBAudio-only DiT, LoRA merged
dramabox-audio-components.safetensors1.9 GBConnector, projection, audio VAE, vocoder
gemma-3-12b-it-bnb-4bit~8 GBText encoder, auto-downloaded

How to run Dramabox locally

Clone the DramaBox repository and either run single generations through src/inference.py or keep a warm TTSServer for repeated calls; weights download on first run. A hosted demo Space is available on Hugging Face if you want to hear it before committing the VRAM.

For chat models to script the dialogue with, see the full catalog of local models.

Dramabox license

The weights ship under the LTX-2 Community License, inherited from the Lightricks base model. Read the license file in the repository before commercial use.

Get the weights from Hugging Face

git clone https://github.com/resemble-ai/DramaBox
python src/inference.py \
    --prompt 'A woman speaks warmly, "Hello, how are you today?"' \
    --voice-sample reference.wav \
    --output output.wav \
    --cfg-scale 2.5 --stg-scale 1.5
from src.inference_server import TTSServer

server = TTSServer(device="cuda")              # downloads weights on first run

server.generate_to_file(
    prompt='A woman speaks warmly, "Hello, how are you today?" '
           'She laughs, "Hahaha, it is so good to see you!"',
    output="output.wav",
    voice_ref="reference.wav",                  # optional, 10+ seconds of target voice
    cfg_scale=2.5,
    stg_scale=1.5,
    seed=42,
)
Desktop



macOS
(M1 or better)
Download
Windows
(x64)
Download
Linux
(x86_64)
Download
Frequently asked questions
Dramabox is Resemble AI's expressive text-to-speech model, an IC-LoRA fine-tune of the LTX-2.3 3.3B audio-only Diffusion Transformer with flow matching, conditioned on Gemma 3 12B text embeddings. The prompt itself controls everything: speaker identity, emotion, delivery, laughs, sighs, breaths, pauses and transitions.
Yes. An optional voice reference of 10 or more seconds clones the target timbre; without one, the model picks a voice that fits the speaker description in your prompt. The reference conditioning window is adjustable from 3 to 30 seconds.
Resemble AI lists about 24 GB of VRAM peak with the warm server. The download is a 6.6 GB audio DiT, 1.9 GB of audio components, and a roughly 8 GB 4-bit Gemma 3 12B text encoder fetched automatically on first run. Warm generation takes about 2.5 seconds per clip on an H100.
The weights are openly available under the LTX-2 Community License, inherited from the Lightricks base model. Read the license file in the repository before commercial use.
Yes, by default. Every output is marked with Resemble Perth, an imperceptible neural watermark that survives MP3 and AAC compression and common edits with close to 100% detection accuracy. It can be disabled with the no-watermark flag.