Edge0-AI/Audio8_TTS

What it solves

Audio8 TTS provides compact, efficient models capable of zero-shot voice cloning. It allows users to generate natural-sounding speech from text while mimicking a specific voice using only a short audio sample, without needing to train a model on that specific voice beforehand.

How it works

The project offers two main model families: a 0.6B parameter model focused on high-quality, multilingual speech synthesis, and a 0.1B parameter model designed for extreme portability and CPU-based inference. Users provide a text prompt and, optionally, a reference audio file to clone a target voice. The system supports standard Python-based inference on CUDA-capable GPUs, as well as optimized deployment via ONNX runtimes and SGLang Omni for scalable, API-compatible serving.

Who it’s for

This project is for developers and researchers who need to integrate text-to-speech capabilities into applications where model size, performance, and the ability to clone voices on-the-fly are critical requirements. It is particularly useful for those needing multilingual support or lightweight deployment on constrained hardware.

Highlights

• Zero-shot voice cloning: Generate speech in a target voice using only a brief reference recording.
• Multilingual support: The 0.6B model supports 23 languages, including English, Chinese, Japanese, German, and many others.
• Flexible deployment: Options range from high-performance GPU inference to CPU-optimized ONNX runtimes for smaller footprints.
• Scalable serving: Includes support for SGLang Omni, providing features like dynamic batching and an OpenAI-compatible API.
• Fine-tuning ready: Provides tools for supervised fine-tuning to further adapt models to specific needs.