Fish Speech touts top TTS benchmarks

This is the kind of speech repo AI developers actually care about: not just a demo model, but a full stack with benchmark ambition and production-minded serving. The big caveat is that despite the “open” positioning, the repo ships under Fish Audio’s research license rather than a standard permissive OSS license.

• –The standout hook is controllability: Fish Speech S2 supports natural-language inline tags for delivery changes like laughter, whispering, and tone shifts.
• –Fish Audio claims best-in-class results on several TTS evals, including lower WER than named closed-source competitors on Seed-TTS benchmarks.
• –The architecture is unusually developer-friendly for deployment because it leans on LLM-style serving optimizations through SGLang, including batching and KV-cache tricks.
• –Rapid voice cloning, multilingual support, and multi-speaker generation make it relevant for agents, character apps, dubbing, and synthetic data workflows.
• –The license matters: teams interested in commercial adoption need to read the Fish Audio Research License carefully before treating this like a normal open-source dependency.