Fish Audio just made S2.1 Pro free for a month.

Here’s everything you need to know about it

• Clones any voice from 10 to 15 seconds of audio.
• ~90ms response, fast enough for real conversation
• 83 languages, one model.
• Word-level control over pronunciation and pauses
• Also has open weighted models

Under the hood it uses a Dual-AR design: a 4B-parameter model works out what to say and how it should feel, and a 400M-parameter model fills in the fine sound detail. That split is why it's both fast and expressive.

It's built from Fish Speech, their open-source project with tens of thousands of GitHub stars, and they still ship open-weight models you can self-host.

Price: roughly 1/6th of ElevenLabs for the same output.

The real pitch isn't a clean 10-second clip. It's a voice that survives a whole conversation, interruptions, corrections, laughter and language switches included.

🧵 1.