Scenema Audio turns voice cloning into stage acting

Anyone who has tried voice cloning knows the pain: feed the model a calm, polite sample, and every single sentence comes out sounding like a customer service rep. If you need raw fury, you have to dig up an angry clip first.

Scenema AI wants to break that constraint with its new open-source model, Scenema Audio.

Extracted from the open-source video generator LTX 2.3, Scenema Audio is an audio diffusion model designed to separate voice identity from emotional performance. Instead of locking vocal dynamics to the source file, it uses text prompts and stage directions to dictate mood, pacing, and breath control.

Giving cold voices wild emotions

The team demonstrated this with a calm, clipped Star Wars dialogue sample. By attaching a stage direction specifying shaking hands and rising disbelief over a winning lottery ticket, the cloned voice transforms entirely, delivering trembling, breathless excitement.

Traditional zero-shot cloning forces you to borrow emotion from the reference audio. Scenema Audio flips this logic: the model first crafts the theatrical performance purely from the prompt, and only then transfers the vocal identity from a few clean seconds of audio. Any voice can express any emotion, even if that person was never recorded in that state.

It handles accent shifts just as easily. In one test, an Australian comedy recording was flipped to an American accent simply by changing the prompt, complete with stage directions for someone giggling so hard they can barely spit out the punchline.

Creators can also drop action tags directly into script text to shift moods mid-sentence, moving from restrained fury to complete collapse. Polyglot cloning works out of the box as well; an English clip of Barack Obama can reliably generate native-sounding speech in Spanish, French, German, Italian, and Swahili.

If you don't want to clone an existing person, you can just describe a character from scratch. A prompt specifying a gravelly, warm voice telling bedtime stories around a dying fire produces a completely synthetic speaker.

Extracted from a video generator

The engineering behind Scenema Audio is surprisingly clever. Instead of building a bespoke text-to-speech engine from scratch, the team extracted the audio diffusion backbone directly from LTX 2.3, noting that top-tier multimodal models are already remarkably adept at standalone audio.

The code, weights, and a desktop GUI are now hosted on GitHub and Hugging Face. Running the quantized version locally requires a GPU with at least 16GB of VRAM and more than 32GB of system RAM.

Why it matters

We think Scenema Audio takes direct aim at the biggest flaw in synthetic audio: the robotic lack of acting.

For indie game developers, filmmakers, and digital storytellers, getting dynamic performances out of synthetic voices usually meant stitching together dozens of mismatched takes or settling for flat narration. Letting creators treat the prompt box like a script with stage directions makes performative voice design practical on consumer hardware.

It also highlights an interesting shift in open-source AI. High-end video models have become so capable that their sub-modules are starting to outclass dedicated single-task tools when decoupled.

The 16GB VRAM barrier will keep casual hobbyists away, and less-represented languages like Swahili still show slight phonetic drift. But for serious creators looking for nuanced audio acting without paying cloud platform fees, Scenema Audio is well worth testing.