Fish Speech S2 Pro: voice cloning and emotion tests

This sponsored review tests Fish Speech's S2 Pro model through the hosted fish.audio interface. The source links the official downloadable S2 Pro weights, but the demonstration does not show a working local installation. The presenter reports an unsuccessful 16 GB Colab attempt and estimates around 24 GB of VRAM for local use. These are his hardware observations, not a tested requirement for every setup.

The presenter calls the release open source. Its official model card specifies the Fish Audio Research License, with restrictions on commercial use; available weights do not mean an unrestricted open-source license.

Voice-cloning examples use roughly 15 seconds of Donald Trump audio and 10 seconds of Elon Musk audio. The presenter compares two outputs with different seeds, then adds laughing and chuckling tags and combines angry and shouting directions. He also describes custom bracketed instructions for delivery beyond preset emotions.

A dialogue example pairs the two cloned voices with separate emotional tags. French, German, Spanish, Arabic, Japanese and Chinese samples use the same English reference voice. The presenter thinks the voice identity remains consistent but notices an underlying English accent and asks native speakers to assess pronunciation.

The review also reports Fish Audio's blind-test results against other speech models. Those vendor figures and the creator's favorable listening impressions are not independent verification, and the competitors are not directly tested here.