cortico-world-gpt-sovits-for-coopanion

Project Overview

This extension by Phantivia is the voice World of the Coopanion extension system: before each pet_say, Coo calls sovits_speak in the same reply, and the extension streams the sentence through local GPT-SoVITS to the sound card. sovits_speak returns only after the first audio block hits the card, so the following pet_say bubble appears in sync with the voice.

Item	Value
Host	Coopanion (desktop pet)
Engine	GPT-SoVITS api_v2
Tiers	GPU with V5 turbo, CPU with v2ProPlus, auto by default
Config	worlds.gpt-sovits

The pet speaks, and the voice is synthesized locally in real time.

Key Features

• Two tiers: with an NVIDIA GPU it uses V5 turbo (first packet 0.5–1.0 s, 3.5 GB VRAM measured); otherwise v2ProPlus on CPU (0.8–1.1 s, 5.8 GB RAM)
• Preset voices: female and male bundled (generated by MOSS-VoiceGenerator, parameters in assets/voices/prompts.json), or upload a 3–10 s custom reference clip
• Three backends: connect an existing api_v2 via serverUrl, point runtimeDir at your own GPT-SoVITS folder, or use the panel's managed install (standalone CPython 3.12 and weights, SHA-256 verified)
• Tools and events: sovits_speak(text) queues synthesis, sovits_hush() stops and clears, and the gpt-sovits.state event broadcasts availability
• Tunable parameters: speed, volume (0–1), output device name, and prompt language (default zh)

Install and Enable

1. Open the GPTSovits Voice page in Coopanion's advanced mode
2. Keep tier: auto — a visible nvidia-smi GPU picks the GPU tier; first launch compiles CUDA kernels (74 s measured, then about 11 s)
3. Warm up with one sentence each in Chinese, English, and Japanese; the service is ready after warmup
4. Without the extension or an un-called sovits_speak, pet_say stays silent and the pet works as usual

Permissions and Scope

• Local service only: the backend runs runtime/python.exe -I -X utf8 -u api_v2.py listening on a free 127.0.0.1 port
• Disk footprint: managed installs place runtimes and weights under runtimes/gpt-sovits/<version>/<tier>/ and models/gpt-sovits/pretrained_models/
• Volume compensation: v2ProPlus output is quiet (about −27 dBFS peak), so playback multiplies by 8 with soft limiting

Notes

• GPU model hates CPU: V5 turbo on CPU takes 3–4 s for the first packet at a 1.2–1.5 real-time factor, so the CPU tier is pinned to v2ProPlus
• Tier changes need a restart: tier and backend settings take effect after restarting the inference service
• Events over snapshots: the voice state in environment prompts is only a prefix-rebuild snapshot; live availability follows the gpt-sovits.state event