Edge0 Releases Audio8-TTS-Realistic, a 0.6B Voice Cloning TTS Model
Summary
- Edge0 released Audio8-TTS-Realistic, an open-weight 0.6B-parameter text-to-speech model supporting zero-shot voice cloning across 23 languages under the Apache 2.0 license.
- The model utilizes a dual-autoregressive architecture combining a Qwen2.5 0.5B slow branch with a 4-layer fast branch predicting codebooks in parallel.
- Training features a three-stage pipeline incorporating supervised fine-tuning and GRPO guided by a learned realism critic to improve natural prosody and rhythm.
- Audio8-TTS-Realistic achieves a 4.86% average error rate on the CV3 benchmark, outperforming several larger open-source models like Qwen3-TTS-1.7B.
- Deployment is supported via PyTorch and Transformers, with an SGLang Omni adapter available for GPU serving with an OpenAI-compatible API.
Why it matters
Audio8-TTS-Realistic provides developers and researchers with a compact, highly capable open-weight voice cloning model that balances low serving costs with competitive multilingual performance.