- https://huggingface.co/Audio8/Audio8-TTS-Preview-0.6b
- 0.6B DualAR TTS (Fish S2 Pro-style): slow AR predicts semantic tokens, fast AR fills codec codebooks. Bundled 44.1 kHz codec, no extra vocoder.
- Zero-shot clone from a short ref clip + exact transcript. 11 languages in this preview (EN, ZH, JA, KO, and several European). Apache 2.0.
- Competitive Seed-TTS EN WER (1.506) vs much larger models. CPU path: ONNX INT4 (~1 GiB). GPU serving: SGLang Omni.
Connections
- part_of Text-to-Speech