Fish Speech S2: hosted voice generation for an AI companion

The creator builds a second AI anime companion, Eno, with a different appearance and personality from his existing character, Riko. He explains the voice conversation pipeline: speech recognition turns spoken input into text, an LLM produces a reply, and text-to-speech converts that reply into audio. The avatar supplies the visual side of the character.

The build starts with shrine maiden concept art, then moves into VRoid Studio to assemble a 3D avatar from presets and purchased outfit and hair assets. For personality, the creator experiments with system prompts intended to produce polite, reserved dialogue. He says he would usually use fine-tuning, but chooses prompting for this build. The video does not identify the LLM or provide a complete installation procedure.

Fish Audio sponsors the voice section. The presenter explicitly identifies its S2 models; this is hosted use of the Fish Speech model family, not a demonstrated download or local inference setup. The creator demonstrates voice cloning and says a 10-second recording can produce a clone in roughly six or seven seconds. He also tries inline expression tags with those models and selects a character voice through the discovery tab. These are demonstrations and timing claims from the speaker, rather than general performance guarantees.

For viewers interested in local AI, the creator describes hardware demands and setup effort as obstacles to running speech synthesis locally. This build uses Fish Audio for the demonstrated voice workflow; it does not establish an offline setup. A basic animation loop and conversation scenes finish the walkthrough.