Skip to content
BackApp Details
Audio8-TTS-Realistic-Air
A compact, open-source (Apache 2.0) 0.6B TTS model delivering ultra-realistic zero-shot voice cloning across 23 languages
Note: Please open in the PC Client. Go to download >>>
Features
Supported Languages
System Requirements
Minimum 8GB RAM. 10GB+ storage recommended. macOS 15+: Supports Intel and M-series chips. Windows 10/11: GPU acceleration (NVIDIA / AMD / Intel Arc discrete GPUs) is recommended, with more than 2GB VRAM required. Running on CPU will be slow, and some older CPUs may fail to generate audio.
Introduction
Audio8_TTS is a lightweight, high-quality open-source text-to-speech (TTS) and voice cloning framework. Its latest mature release is Edge0/Audio8-TTS-Realistic-Air (0.6B).
Connection to Preview-0.6b:
Both share the exact same 601M parameter scale, the core DualAR architecture, the bundled 44.1 kHz neural audio codec, and the same Hugging Face Transformers inference interface.
Key Differences from Preview-0.6b:
- Expanded Multilingual Support: While the Preview version supported 11 languages, Realistic-Air expands coverage to 23 languages (adding Arabic, Finnish, Hindi, Indonesian, Malay, Norwegian, Portuguese, Russian, Swedish, Thai, Turkish, and Vietnamese).
- Substantially Improved Realism: Realistic-Air introduces advanced GRPO reinforcement learning guided by a specialized "Realism Critic," significantly refining prosody, natural breathing pauses, and rhythm to overcome the robotic artifacts present in the early preview.
Key Features & Product Highlights
- Zero-Shot Voice Cloning: Instantly replicate any voice and generate custom speech for new text by providing just a short reference audio clip and its transcript.
- Direct Text-to-Speech: Functions out-of-the-box as a standard, high-quality TTS engine without needing a voice reference.
- Lightweight & Commercial-Friendly: Boasting a compact 0.6B footprint with minimal VRAM/RAM demands, both weights and code are released under the permissive Apache 2.0 license for free commercial use and fine-tuning.
Development Team & Target Use Cases
- Developer: Developed and open-sourced by Edge0 (Edge0-AI).
- Use Cases:
- On-Device / Edge Deployment: Ideal for local execution on personal computers or resource-constrained servers.
- Content Creation & Gaming: Rapidly generate custom voiceovers for videos, audiobooks, and NPC dialogues.
- Real-Time Voice Agents: Features built-in support for OpenAI-compatible API endpoints (
/v1/audio/speech) for seamless conversational integration.
Underlying Architecture & Technology
- Dual-Autoregressive (DualAR) Design:
- Slow AR Backbone: Built on Qwen2.5-0.5B to predict frame-level semantic tokens.
- Fast AR Branch: A 4-layer shallow network that predicts remaining audio codebooks in parallel.
- Integrated 44.1 kHz Neural Audio Codec: Eliminates external vocoder dependencies by handling waveform decoding natively.
- Realism-Guided Reinforcement Learning: Utilizes a learned "Realism Critic" during training to align synthesized timing and natural phrasing with human speech patterns.
LicenseApache-2.0