DramaBox, Voice Model Released

Resemble AI has released DramaBox.

This model is an IC-LoRA tuning layered on top of LTX-2.3 3.3B audio-only, a prompt-based TTS that controls speaker identity, emotion, tone, laughter, sighs, pauses, and transitions using only prompts. Optionally, a 10+ second voice reference can be provided to clone the target timbre, and text embeddings use Gemma 3 12B.

• A GitHub, Hugging Face model card, and Hugging Face Space were released together.
• The inference weights are dramabox-dit-v1.safetensors at 6.6GB, and the audio component is 1.9GB.
• On a warm server, generation time is presented as about 2.5 seconds, with peak VRAM at about 24GB.
• Outputs are watermarked by default with Resemble Perth's neural network watermark, which can be disabled with --no-watermark.
• The base model is Lightricks/LTX-2.3, and the license is the LTX-2 Community License.