Fish Audio S2 open-sources, tops all TTS benchmarks

S2 is the most credible open-source TTS challenger to ElevenLabs yet — and the inline freeform emotion tagging system (`[whispers sweetly]`, `[laughing nervously]`) is a genuine UX leap over rigid fixed-schema alternatives.

• –Beats every evaluated closed-source model on Seed-TTS WER (0.54% Chinese, 0.99% English) and EmergentTTS-Eval (81.88% win rate vs. gpt-4o-mini-tts)
• –Dual-AR architecture (Qwen3-4B slow AR + 400M fast AR) maps directly onto the LLM serving stack — SGLang integration means production streaming without custom infra
• –Zero-shot voice cloning from 10–30 second clips with 86.4% RadixAttention cache hit rates makes it practical for high-volume API workloads
• –Caveat: the Research License bars commercial use without a separate Fish Audio agreement — not truly permissive open-source, which disappointed parts of the community
• –Hardware ceiling is real: 12 GB VRAM minimum, real-time RTF only on H200-class GPUs; consumer setups will struggle