Reference-Conditioned Multi-Speaker Audio Model — Anonymous Review Release
Model weights accompanying an anonymous submission under peer review. This is an audio-only, reference-conditioned text-to-audio flow-matching model built on the LTX-2 architecture. Given a natural-language prompt and one or more reference-audio clips, it generates a multi-speaker audio scene in which each reference voice is bound to its speaker through the text.
Files
| file | contents |
|---|---|
model.safetensors |
the transformer weights + the Gemma text-projection head |
audio_vae.safetensors |
the audio VAE + vocoder (bundled, so the full LTX-2 model is not required) |
Total ≈ 8.5 GB.
Required external dependency: Gemma
The text encoder is Google's gemma-3-12b-it, which is not included here. Obtain
it separately from its official source under its own license and point the code's
--gemma-root at the local copy.
Usage
Use the accompanying anonymous code repository. In brief:
python examples/generate.py --gemma-root /path/to/gemma-3-12b-it \
--ckpt-dir /path/to/this/repo --out out.wav
License
These weights are released under the LTX-2 Community License Agreement of the base
architecture; the full license text is embedded in the model.safetensors header
metadata (__metadata__.license) and reproduced in LICENSE. You must also comply with
the Gemma license for the separately obtained text encoder.
Copyright (c) 2026 Anonymous Authors, for this review release.