ScenA is a reference-driven model designed for generating multi-speaker audio scenes from text prompts. The tool produces dialogue or conversation audio by conditioning the output on one or more reference audio clips, which determine the voices of the speakers. This approach allows for the creation of audio scenes where each speaker's voice is guided by provided audio samples, making it possible to synthesize complex, multi-speaker interactions from written text.
The model operates as an audio-only, reference-conditioned text-to-audio system. It uses a flow-matching DiT architecture built on LTX-2, featuring approximately 4 billion parameters across 48 layers. ScenA’s architecture incorporates a transformer and Gemma text-projection weights, as well as an audio VAE (variational autoencoder) for encoding and decoding audio, including vocoding capabilities. These components enable the model to generate high-fidelity, multi-speaker audio scenes based on the specified text and reference voices.
safetensors provided for download. The tool is released under the ltx-2-community-license. It is positioned for use in English language audio generation tasks, with a focus on applications that require reference-based speaker conditioning and dialogue synthesis.
Developed in association with Lightricks and Tel Aviv University, ScenA is intended for researchers and developers interested in advanced audio generation, speech synthesis, and multi-speaker scene creation from text prompts. Documentation, project details, and code are accessible through linked project and GitHub pages referenced in the model card.
Scena is a Voice, TTS & speech product. It focuses on generating realistic multi-speaker audio scenes from text prompts and reference audio using open-source AI. Scena is an open-source project aimed at AI audio researchers and developers. The project is open source (Open Source). The product ships for the web, the command line, and API, and it can be self-hosted.
suryatmodulus builds and maintains Scena, and the product first shipped in 2026. The project is developed in the open on GitHub with 17 stars and 10 commits in the last 90 days. Among its 5 catalogued features are audio scene generation, multi-speaker support, and reference conditioning.
Latest indexed changes and source events
suryatmodulus/scena verified by the PulseGate indexer
Other apps tracked under the same category.
