Resemble AI releases DramaBox, an open-source TTS model for directed voice performances

Resemble AI released DramaBox, an open-source text-to-speech model based on LTX-2.3 and Gemma 3 12B that interprets stage directions for cinematic voices.

Resemble AI releases DramaBox, a text-to-speech model designed for narrative and cinematic production, open source. The system operates like a film script: dialogues are written within quotation marks, and stage directions placed outside quotation marks (sighs, pauses, laughs, whispers, intonation variations) are interpreted by the model without being spoken. The output is 48 kHz stereo, studio quality, and allows for both zero-shot casting by description (age, accent, mood) and voice cloning from about ten seconds of reference audio. Each file includes Resemble's inaudible PerTh watermark by default, which is resistant to compression and tampering, but can be disabled depending on the use case.

Under the hood, DramaBox is an IC-LoRA fine-tune of Lightricks' LTX-2.3 audio model (3.3 billion parameters), conditioned by Gemma 3 12B. Inference is reported at approximately 2.5 seconds on an H100, with a peak of 24 GB of VRAM. The model is available on Hugging Face and GitHub, along with a Gradio interface, an inference server, and the ability to fine-tune custom LoRAs to add personalized voices or styles.