Description
Audio8-TTS-Realistic-Air is an advanced text-to-speech (TTS) model developed by Edge0, available on Hugging Face. Building upon the Audio8-TTS-Preview-0.6b, this release significantly improves speech realism, delivering more natural prosody and expressive delivery, while maintaining a compact 0.6 billion parameter architecture.
Read More
The model boasts broader multilingual coverage, supporting 23 languages. This includes the 11 languages from the Preview version (Cantonese, Chinese, Dutch, English, French, German, Italian, Japanese, Korean, Polish, Spanish) and 12 new additions (Arabic, Finnish, Hindi, Indonesian, Malay, Norwegian, Portuguese, Russian, Swedish, Thai, Turkish, Vietnamese). It utilizes the same Transformers inference interface as its predecessor, making it accessible for developers.
Audio8-TTS-Realistic-Air includes a complete model checkpoint, a 44.1 kHz neural audio codec, tokenizer, processor, and custom Transformers code. It supports flexible generation, offering both zero-shot voice cloning from a reference recording and transcript, as well as speech generation without a reference. Users can experiment with these capabilities in an interactive Playground hosted on Hugging Face ZeroGPU.
For high-throughput applications, the model features GPU serving via an SGLang Omni adapter. This adapter provides paged attention, dynamic batching, and an OpenAI-compatible speech API, ensuring efficient and scalable deployment. The model's architecture is a DualAR design, inspired by Fish Audio S2 Pro, with a slow AR transformer predicting semantic tokens and a fast AR transformer predicting codec codebooks.
Installation requires Python 3.10 or newer and a CUDA-capable GPU. The model has been smoke-tested with PyTorch 2.8.0 and Transformers 4.57.3 on an NVIDIA H20, covering various language generations and voice cloning. The code and model weights are released under the Apache License 2.0, promoting open access and responsible use. While quality varies by language, the model represents a significant step forward in realistic and multilingual text-to-speech synthesis.
- CategoryAI Models & LLMs
- Pricing ModelOpen source
- DeploymentCloud
- PlatformsWeb
Audio8-TTS-Realistic-Air Highlights
Improved speech realism with natural prosody and expressive delivery
Broader multilingual coverage supporting 23 languages
Compact model architecture with 601 million main-model parameters
Zero-shot voice cloning from reference audio and transcript
Speech generation without a reference recording
GPU serving with SGLang Omni adapter for high throughput
OpenAI-compatible /v1/audio/speech API support
DualAR architecture for advanced audio processing
Bundled 44.1 kHz neural audio codec
Includes tokenizer, processor, and custom Transformers code
Supports dynamic batching and paged attention
Apache License 2.0 for code and model weights
Interactive Playground for testing speech generation and voice cloning
Getting Started with Audio8-TTS-Realistic-Air
Access Model: Navigate to the Edge0/Audio8-TTS-Realistic-Air page on Hugging Face.
Set Up Environment: Ensure Python 3.10+ and a CUDA-capable GPU are installed.
Load Model: Load the model using custom Transformers code with `trust_remote_code=True`.
Generate Speech: Use the processor to prepare inputs for speech generation.
Perform Voice Cloning: Provide a clean reference recording and its exact transcript for zero-shot voice cloning.
Run Inference: Move inputs to the model device and execute generation and decoding steps.
Deploy for Serving: Utilize the SGLang Omni adapter for high-throughput GPU serving with an OpenAI-compatible API.
Audio8-TTS-Realistic-Air's Use Cases
- Realistic Speech Generation
- Multilingual Content Creation
- Zero-Shot Voice Cloning
- Interactive Voice Applications
- High-Throughput Audio Serving
- Research and Development








