declare-lab/tango

Leverages a frozen Flan-T5 instruction-tuned LLM for text encoding paired with a UNet-based latent diffusion model, enabling realistic audio synthesis across diverse sound categories from minimal training data. Tango 2 adds alignment via Direct Preference Optimization (DPO) on the Audio-Alpaca preference dataset for improved output quality. Integrates with Hugging Face (model hosting, Accelerate for multi-GPU training) and provides inference via simple Python API with configurable sampling steps for quality-speed tradeoffs.

1,233 stars. No commits in the last 6 months.