High-Logic/Genie-TTS

██████╗  ███████╗███╗   ██╗██╗███████╗
██╔════╝ ██╔════╝████╗  ██║██║██╔════╝
██║  ███╗█████╗  ██╔██╗ ██║██║█████╗  
██║   ██║██╔══╝  ██║╚██╗██║██║██╔══╝  
╚██████╔╝███████╗██║ ╚████║██║███████╗
 ╚═════╝ ╚══════╝╚═╝  ╚═══╝╚═╝╚══════╝

---

GENIE is a lightweight inference engine built on the open-source TTS project GPT-SoVITS. It integrates TTS inference, ONNX model conversion, API server, and other core features, aiming to provide ultimate performance and convenience.

• ✅ Supported Model Version: GPT-SoVITS V2, V2ProPlus
• ✅ Supported Language: Japanese, English, Chinese, Korean
• ✅ Supported Python Version: >= 3.10

🎬 Demo Video

• ➡️ Watch the demo video (Chinese)

🚀 Performance Advantages

GENIE optimizes the original model for outstanding CPU performance.

| Feature | 🔮 GENIE | Official PyTorch Model | Official ONNX Model | |:----------------------------|:-----------:|:----------------------:|:-------------------:| | First Inference Latency | 1.13s | 1.35s | 3.57s | | Runtime Size | \~200MB | \~several GB | Similar to GENIE | | Model Size | \~230MB | Similar to GENIE | \~750MB |

📝 Latency Test Info: All latency data is based on a test set of 100 Japanese sentences (\~20 characters each),

averaged. Tested on CPU i7-13620H.

---

🏁 QuickStart

⚠️ Important: It is recommended to run GENIE in Administrator mode to avoid potential performance degradation.

📦 Installation

Install via pip:

pip install genie-tts

📥 Pretrained Models

When running GENIE for the first time, it requires downloading resource files (~391MB). You can follow the library's prompts to download them automatically.

Alternatively, you can manually download the files

from HuggingFace

and place them in a local folder. Then set the GENIE_DATA_DIR environment variable before importing the library:

import os
Set the path to your manually downloaded resource files
Note: Do this BEFORE importing genie_tts
os.environ["GENIE_DATA_DIR"] = r"C:\path\to\your\GenieData"
import genie_tts as genie
The library will now load resources from the specified directory

If you want the optional Chinese RoBERTa text features used only for Chinese inference to improve Chinese prosody, you can also download them with:

import genie_tts as genie
Download only the optional Chinese RoBERTa assets
genie.download_roberta_data()
Or use the built-in full resource download flow,
which now also downloads the optional Chinese RoBERTa assets
genie.download_genie_data()

These RoBERTa features are intended only for the Chinese path to improve Chinese prosody. They should not be used for non-Chinese inference (Japanese / English / Korean).

⚡️ Quick Tryout

No GPT-SoVITS model yet? No problem! GENIE includes several predefined speaker characters you can use immediately — for example:

• Mika (聖園ミカ) — Blue Archive (Japanese)
• ThirtySeven (37) — Reverse: 1999 (English)
• Feibi (菲比) — Wuthering Waves (Chinese)

Try it out with the example below:

import genie_tts as genie
import time
Automatically downloads required files on first run
genie.load_predefined_character('mika')
genie.tts(
    character_name='mika',
    text='どうしようかな……やっぱりやりたいかも……！',
    play=True,  # Play the generated audio directly
)
genie.wait_for_playback_done()  # Ensure audio playback completes

🎤 TTS Best Practices

A simple TTS inference example:

import genie_tts as genie
Step 1: Load character voice model
genie.load_character(
    character_name='<CHARACTER_NAME>',  # Replace with your character name
    onnx_model_dir=r"<PATH_TO_CHARACTER_ONNX_MODEL_DIR>",  # Folder containing ONNX model
    language='<LANGUAGE_CODE>',  # Replace with language code, e.g., 'en', 'zh', 'jp', 'kr'
)
Step 2: Set reference audio (for emotion and intonation cloning)
genie.set_reference_audio(
    character_name='<CHARACTER_NAME>',  # Must match loaded character name
    audio_path=r"<PATH_TO_REFERENCE_AUDIO>",  # Path to reference audio
    audio_text="<REFERENCE_AUDIO_TEXT>",  # Corresponding text
)
Step 3: Run TTS inference and generate audio
genie.tts(
    character_name='<CHARACTER_NAME>',  # Must match loaded character
    text="<TEXT_TO_SYNTHESIZE>",  # Text to synthesize
    play=True,  # Play audio directly
    save_path="<OUTPUT_AUDIO_PATH>",  # Output audio file path
)
genie.wait_for_playback_done()  # Ensure audio playback completes
print("🎉 Audio generation complete!")

---

🔧 Model Conversion

To convert original GPT-SoVITS models for GENIE, ensure torch is installed:

pip install torch

Use the built-in conversion tool:

Tip: convert_to_onnx currently supports V2 and V2ProPlus models.

import genie_tts as genie
genie.convert_to_onnx(
    torch_pth_path=r"",  # Replace with your .pth file
    torch_ckpt_path=r"",  # Replace with your .ckpt file
    output_dir=r""  # Directory to save ONNX model
)

---

🌐 Launch FastAPI Server

GENIE includes a lightweight FastAPI server:

import genie_tts as genie
Start server
genie.start_server(
    host="0.0.0.0",  # Host address
    port=8000,  # Port
    workers=1  # Number of workers
)

For request formats and API details, see our API Server Tutorial.

---

📝 Roadmap

• [x] 🌐 Language Expansion
• [x] Add support for Chinese and English.
• [x] 🚀 Model Compatibility
• [x] Support for V2ProPlus.
• [ ] Support for V3, V4, and more.
• [x] 📦 Easy Deployment
• [ ] Release Official Docker images.
• [x] Provide out-of-the-box Windows bundles.