scenemaai/scenema-audio

Scenema Audio is a Python tool for generating highly expressive speech audio from text, with the ability to clone any voice from a short audio sample. Unlike standard text-to-speech systems that simply pronounce words, this one is designed to perform them, capturing emotional delivery, pacing, breath, and shifts in tone within a single generation. The prompting system uses an XML format where you describe the voice in natural language, write the speech, and insert action tags to direct emotional and physical delivery between sentences. For example, you can instruct the model that the speaker "voice tightens, swallows, fighting to stay composed" before a specific line, and it reflects that in the audio. A scene attribute adds environmental background sounds like rain or crowd noise. For voice cloning, you provide 10 to 20 seconds of reference audio and the model transfers that speaker's vocal identity onto whatever emotional performance the prompt specifies, even if the reference speaker was never recorded in that emotional state. The system runs locally using Docker on an NVIDIA GPU with at least 16 GB of video memory, downloading around 38 GB of model weights on first launch. A web interface built with Gradio is included for experimenting without writing code, and the tool exposes a REST API for integration into other software.