Resemble AI gives TTS an acting director with DramaBox

Anyone who has ever built something with text-to-speech knows the exact wall you hit. The words are right, the pronunciation is flawless, but the delivery is completely lifeless. Flat as cardboard.

Historically, the fix was fiddling with emotion sliders, tweaking pitch parameters, or guessing which arcane markup tag made an AI voice sound sad without breaking. But in the real world, no movie director tells an actor to increase emotional intensity by twenty percent.

Resemble AI tackled that disconnect head-on by launching DramaBox, an expressive voice model designed to make speech generation feel like directing a scene.

Directing audio like a movie script

DramaBox ditches parameter tuning for a format anyone can understand: a screenplay.

The prompt splits cleanly into two distinct zones. Everything inside double quotation marks is dialogue meant to be spoken out loud. Everything outside double quotes is treated as acting direction and scene context.

Write outside the quotes that a regal character speaks with cold fury and sighs deeply, and the model inhales, lets out an audible breath of frustration, and drops into a measured, venomous rasp. Instruct it that someone is laughing so hard they are gasping for air, and it generates genuine giggles, snorts, and breath-catching pauses right in the middle of lines.

It also handles voice cloning out of the box. Feed it a 10-second reference audio clip and it will mirror that person's vocal timbre while obeying whatever drama you write in the prompt. If you skip the reference clip, it simply invents an original voice that matches the scene description.

Heavy lifting under the hood

Under the hood, DramaBox is built as a 3.3B-parameter audio diffusion transformer paired with a 4-bit quantized Gemma 3 12B text encoder, sharing architectural lineage with video models like LTX 2.3.

It outputs 48kHz stereo audio in AAC or WAV format. On a warm H100 GPU, generating a full performance clip takes roughly 2.5 seconds.

Resemble AI made the model weights downloadable for local inference alongside setup guides. But running it locally requires serious iron. While the audio model alone is 7GB, pairing it with the 8GB text encoder and auxiliary components pushes peak VRAM usage to 24GB. While 16GB might limp along, an enthusiast GPU is practically required.

There is also a strong nod toward enterprise compliance. Every audio file embeds Resemble Watermarker directly into speech-relevant frequencies. The watermark is invisible to human ears but survives MP3 compression and re-encoding with near 100% detection accuracy, aiming directly at upcoming requirements under the EU AI Act. Open-source machine learning engineer Joshua Lochner celebrated the release, highlighting that downloads across the Resemble TTS family on Hugging Face have now surpassed 10 million.

Why it matters

In our view, DramaBox hits the mark because it replaces robotic toggles with natural human storytelling.

Directing an AI voice should never feel like calibrating audio equipment. By borrowing screenplay conventions, Resemble makes high-stakes vocal drama accessible to indie game developers, animators, and audiobook producers who previously had to manually stitch together dozens of takes.

The 24GB VRAM ceiling means this is not running on an everyday laptop, which remains a real barrier for casual hobbyists wanting to tinker at home.

Still, pairing expressive open weights with a robust forensic watermark sets a sharp precedent for the industry. As incoming regulations scrutinize synthetic media, models that ignore provenance will become legal liabilities. Proving that an AI voice can act while staying compliant is exactly what will push synthetic audio into prime time.