Neiroha CosyVoice3

This page covers the Neiroha CosyVoice3 local backend. Extract the portable Release package, or place the source repository anywhere; <backend-root> refers to that directory.

It provides FastAPI, Gradio Admin, TOML model presets, TOML voice sets, an OpenAI-compatible TTS API, and native /api/cosyvoice routes. The current version uses an independent CosyVoice3 backend, with official FunAudioLLM/CosyVoice as a submodule.

The backend Admin shows API state, voice sets, model presets, clone configuration, and logs.

Capability Summary

Dimension	Current Notes
Recommended version	Default is Fun-CosyVoice3-0.5B, output sample rate 24 kHz.
Languages	Official model card lists 9 common languages: Chinese, English, Japanese, Korean, German, Spanish, French, Italian, Russian.
Dialects / accents	Official model card mentions 18+ Chinese dialects / accents, including Guangdong, Minnan, Sichuan, Northeast, Shanxi / Shaanxi, Shanghai, Tianjin, Shandong, Ningxia, and Gansu.
Cross-language output	Supports multilingual / cross-lingual zero-shot voice cloning. Target language should still stay within the official 9 languages.
prompt clone	prompt-clone / zero_shot needs reference audio and prompt text matching that audio.
cross-lingual clone	cross-lingual-clone needs reference audio only; target text decides output language.
instruct clone	instruct-clone needs reference audio and instruction for speed, emotion, dialect, volume, and similar controls.
Official speed reference	Official model card emphasizes bi-streaming and latency down to around 150 ms, but does not provide one unified PyTorch RTF table.
Boundaries	Rare words, tongue twisters, and specialized terms can be unstable; emotion control depends heavily on text semantics.

Default Addresses

Service	Default Address	Purpose
FastAPI	http://127.0.0.1:9880	Neiroha provider connects here.
Admin	http://127.0.0.1:7880	Manage voice sets, clone config, model presets, downloads, and logs.

CosyVoice3 and GPT-SoVITS both default to API port 9880. Change one backend port in configs/server.toml, or use the random port selected by the launcher.

Install

Download Windows portable packages from Neiroha-Cosyvoice V1.0.0 Release. The current package is built for NVIDIA GPU / CUDA environments and mainly targets RTX 30 / 40 / 50 series users. Current split archive names are neiroha-cosyvoice3-portable.7z.001 through .006; if GitHub downloads are unstable, use the Baidu Netdisk mirror from the Release body. Put all six parts in the same directory and extract from .001.

Source or development environment:

pixi install
pixi run submodule-init
pixi run install

pixi run install downloads the CosyVoice3 model to:

models/Fun-CosyVoice3-0.5B

Optional resources:

pixi run install-wetext
pixi run install-ttsfrd

Start

Portable Release:

.\start_portable.bat

Source environment:

pixi run serve

Common tasks:

Command	Purpose
pixi run serve	Start according to configs/server.toml [startup].surface, default API + Admin
pixi run api	Start FastAPI only
pixi run admin	Start Gradio Admin only and connect to an existing FastAPI
pixi run smoke	Check /health, /v1/models, /v1/audio/voices, and synthesis
pixi run clone-smoke	Check clone flows
pixi run test	Run backend tests
pixi run launcher-help	Show launcher arguments

Default [startup].preload_model = true, so first startup loads the model.

Connect Neiroha

1. Open Providers.
2. Create a provider, adapter type CosyVoice Native.
3. Set Base URL to http://127.0.0.1:9880, or the actual logged address.
4. Leave API Key empty if local auth is disabled.
5. Click Fetch All.
6. Confirm prompt-clone, cross-lingual-clone, and instruct-clone.
7. Enable the provider and click Health Check.

Android emulator host URL:

http://10.0.2.2:9880

Character Setup

Goal	Setting
Zero-shot prompt clone	Select prompt-clone, provide reference audio and prompt text.
Cross-language clone	Select cross-lingual-clone, provide reference audio and target-language text.
Instruction control	Select instruct-clone, write voice requirements in instruction.
Custom reusable voice	Upload reference audio in Admin clone config and save a new voice.

Use clean short reference audio without background music. prompt_clone requires prompt text; instruct requires instruction.

API Prefix

OpenAI-compatible routes:

Method	Path	Purpose
GET	/health	Health check
GET	/v1/models	List voice sets
GET	/v1/audio/voices	List voice profiles
POST	/v1/audio/speech	Synthesize with registered voice

Native prefix is /api/cosyvoice. Legacy /cosyvoice/* and /cosyvoice3/* remain for compatibility; new integrations should prefer /api/cosyvoice.

Sources

• Fun-CosyVoice3-0.5B-2512 model card
• FunAudioLLM/CosyVoice examples
• CosyVoice3 paper