Fish Speech S2 Pro open-sources 15k inline tags

Fish Speech S2 Pro is a direct shot at ElevenLabs, offering production-grade latency and deep prosodic control that was previously locked behind proprietary APIs.

• –Dual-AR architecture (4B Slow AR + 400M Fast AR) balances linguistic structure with high-fidelity acoustic detail for more natural phrasing
• –Massive library of 15,000+ inline tags like [whisper] and [sigh] allows for granular emotional directing without external conditioning models
• –Optimized for sub-150ms latency on H100/H200 hardware, making it viable for real-time conversational agents and interactive gaming
• –Multilingual support for 80+ languages trained on 10M+ hours of audio puts it in the top tier of open-weights TTS models
• –Early user reports suggest a learning curve for local hardware optimization, but the underlying model quality is a significant leap for the open-source audio ecosystem