Neiroha GPT-SoVITS Local Launcher

Standalone GPT-SoVITS backend for Neiroha. This repository owns its own Pixi environment, FastAPI server, Gradio Admin, launcher scripts, and configuration semantics. It does not share a Python environment with other TTS engines.

Chinese documentation: README_zh.md

The current shape is intentionally small:

• OpenAI-style TTS API: /v1/models, /v1/audio/voices, /v1/audio/speech
• model means a voice set, not a low-level weight model
• voice means a concrete voice profile inside the selected voice set
• low-level GPT-SoVITS runtime weights live in model presets
• only one default voice is exposed: genshin-keqing
• multi-role packs and shared community weights are not bundled as bulk download actions; download them yourself and register the weights in Admin

Layout

configs/
  server.toml
  model-presets/
    default.toml
  voice-sets/
    default.toml
runtime/
  voices/
    genshin-keqing/
      voice.toml
      reference.wav
  logs/
  outputs/
models/
  pretrained/

Neiroha-owned configuration is TOML. The only YAML file is the generated GPT-SoVITS runtime cache at runtime/cache/tts_infer.yaml, because upstream GPT-SoVITS expects that shape.

configs/voice-sets/default.toml maps to model=default in the Neiroha / OpenAI-compatible API. runtime/voices/genshin-keqing/voice.toml maps to voice=genshin-keqing. configs/model-presets/default.toml is the low-level GPT-SoVITS runtime preset.

Install

pixi install
pixi run install
pixi run install-sample-voice

install initializes the submodule, installs upstream GPT-SoVITS Python dependencies, and downloads the official pretrained assets. install-sample-voice downloads only one default sample reference voice; it does not download trained multi-character packs.

Run

start_portable.bat serve

Or use Pixi tasks:

pixi run serve
pixi run api
pixi run admin
pixi run test
pixi run smoke

Default ports come from configs/server.toml:

FastAPI  http://127.0.0.1:9880
Admin    http://127.0.0.1:7860

serve reads configs/server.toml and uses [startup].surface plus [startup].preload_model. Pixi tasks no longer hard-code host, port, surface combinations, or preload policy. If a configured port is busy or Windows refuses the bind, the launcher picks an available random port and prints the actual FastAPI and Gradio Admin URLs in the terminal and runtime/logs/backend.log.

admin starts only the Gradio Admin and connects to an existing FastAPI server. Set [startup].surface = "both" and run pixi run serve to start FastAPI and Gradio Admin together. Gradio is not mounted into FastAPI.

Admin Language

Gradio does not automatically translate custom labels, but this Admin supports Chinese and English labels at launch time. Set:

# configs/server.toml
[ui]
default_language = "en" # zh | en

Or set an environment variable:

$env:NEIROHA_GPT_SOVITS_UI_LANG="en"

Then restart Admin.

Logs

The Admin Logs tab shows runtime/logs/backend.log, newest entries first, with automatic refresh. Each top-level launcher start rotates the previous file to runtime/logs/backend.previous.log and starts a fresh backend.log, so normal use does not accumulate one endless log file. Download tasks still write fresh per-run files:

runtime/logs/admin-download.out.log
runtime/logs/admin-download.err.log

API

List voice sets:

curl.exe http://127.0.0.1:9880/v1/models

List voices:

curl.exe http://127.0.0.1:9880/v1/audio/voices

Synthesize speech:

curl.exe http://127.0.0.1:9880/v1/audio/speech `
  -H "Content-Type: application/json" `
  -d '{ "model":"default", "voice":"genshin-keqing", "input":"Hello, this is a voice cloning test.", "response_format":"wav" }' `
  --output speech.wav

Responses include X-Neiroha-Output-Path, X-Neiroha-Audio-Seconds, X-Neiroha-Elapsed-Seconds, and X-Neiroha-RTF headers.

Add Voices

Use the Admin Voice Config tab to upload reference audio, enter the matching prompt text, set a voice id/name, and save it.

By default the new voice uses the selected model preset. For community-trained GPT-SoVITS v2Pro/v2ProPlus weights, either add a reusable preset in the Model Presets tab, or expand the per-voice override section and fill the .ckpt and .pth paths directly.

Manual voice profile path:

runtime/voices/my_voice/voice.toml

Then add my_voice to:

configs/voice-sets/default.toml

Example:

schema_version = 1
id = "my_voice"
name = "My Voice"
mode = "prompt_clone"
model_preset = "v2proplus-clone"
reference_audio = "runtime/voices/my_voice/reference.wav"
prompt_audio = ""
prompt_text = "Transcript matching the reference audio."
text_lang = "zh"
prompt_lang = "zh"
instruction = ""
speed = 1.0
gpt_weights_path = "models/voices/my-trained/GPT_xxx.ckpt"
sovits_weights_path = "models/voices/my-trained/SV_xxx.pth"

[engine_options]

gpt_weights_path and sovits_weights_path are optional per-voice overrides. If omitted, the voice uses the weights from model_preset.