A powerful, local text-to-speech system powered by Llasa-tts models with a modern, interactive dashboard interface. This project supports both 3B and 8B models and features advanced caching, GPU optimizations, and customizable generation parameters.
Llasa TTS Dashboard transforms the traditional TTS pipeline into a user-friendly, robust application. Key improvements include:
- Revamped Dashboard UI: A sleek, two-panel interface for synthesis and history, built with Gradio.
- Multi-Model Support: Easily switch between 3B and 8B models.
- Enhanced Inference Options: Generate speech from plain text or using a reference audio prompt.
- Optimized Performance: Model caching, efficient CUDA initialization, and precise GPU memory management ensure smooth performance.
- Advanced Generation Controls: Adjust parameters such as max length, temperature, top-p, and set a random seed for reproducibility.
- Comprehensive Environment Setup: Pre-configured for optimal HuggingFace cache management and progress reporting.
- VRAM Requirements:
- 8.5 GB VRAM: When loading Whisper Large Turbo alongside the LLM in 4-bit.
- 6.5 GB VRAM: When running without Whisper and using the LLM in 4-bit.
- GPU: An NVIDIA GPU with CUDA support is required.
Clone the repository, install the dependencies, and run the application using the following commands:
git clone https://github.com/nivibilla/local-llasa-tts.git
cd local-llasa-tts
pip install -r requirements_native_hf.txt
pip install -r requirements_base.txt
python hf_app.pyIf you are unable to run the application locally, try our Colab Notebook:
- Long Text Inference with VLLM and Chunking:
See the llasa_vllm_longtext_inference.ipynb notebook for details on handling long text inputs.
- Inspiration & Contributions:
- Based on the original LLaSA training repository.
- Grateful to mrfakename for the Gradio demo code.
Enjoy the new dashboard and the enhanced text-to-speech generation experience!