LunaVox is a lightweight inference engine based on the open-source TTS project GPT-SoVITS. It bundles speech synthesis, ONNX model conversion, an API server, and other conveniences to deliver faster deployment and better ergonomics.

• Supported model versions: GPT-SoVITS V2, GPT-SoVITS V2 Pro Plus
• Supported languages: Japanese, Chinese, English

LunaVox preserves the core GPT-SoVITS inference pipeline: multilingual front-ends (e.g., Open JTalk) convert text to phonemes → HuBERT extracts reference audio features → a three-stage T2S stack (Encoder / First-Stage Decoder / Stage Decoder) produces speech tokens → the VITS vocoder renders the final waveform. All of these components—including the Chinese HuBERT and speaker vector models—are provided as ONNX graphs and paired with caching so that pure ONNX Runtime inference remains fast and resource friendly.

Install via pip:

pip install lunavox-tts

Note: Installing pyopenjtalk may fail because it ships native extensions without prebuilt wheels. On Windows you must install the Visual Studio Build Tools and enable the “Desktop development with C++” workload.

All demo scripts live under Tutorial/ and will first call Tutorial/data_setup.py to download any missing assets plus set HUBERT_MODEL_PATH / OPEN_JTALK_DICT_DIR for you. Run them from the repo root:

python Tutorial/v2_quick_tryout/quick_tryout_en.py  # English prompt + output
python Tutorial/v2_quick_tryout/quick_tryout_zh.py  # Chinese prompt + output
python Tutorial/v2_quick_tryout/quick_tryout_ja.py  # Japanese prompt + output

python Tutorial/v2_pro_plus_quick_tryout/quick_tryout_v2proplus_en.py
python Tutorial/v2_pro_plus_quick_tryout/quick_tryout_v2proplus_zh.py
python Tutorial/v2_pro_plus_quick_tryout/quick_tryout_v2proplus_ja.py

The v2 Pro Plus scripts need the ERes2NetV2 speaker embedding model exported to Data/sv/eres2netv2.onnx; follow the documentation’s export steps before running them.

For users in mainland China we recommend downloading the required models and dictionaries manually and placing them inside the root Data directory.

After downloading, point to the assets with environment variables (os.environ).

• Chinese text pipeline (lunavox_tts.Chinese.ZhBert)
  Install with pip install "lunavox-tts[zh]" to pull in torch and transformers. Without the extra, Chinese inputs fall back to zero BERT embeddings while Japanese/English inference keeps working.
• Model conversion utilities (lunavox.convert_to_onnx)
  Install with pip install "lunavox-tts[convert]" to enable the PyTorch-based converter.

Example for multilingual synthesis:

import os

# Optional: point to the Chinese HuBERT model. If omitted, the script will try to download it from Hugging Face.
os.environ['HUBERT_MODEL_PATH'] = r"C:\path\to\your\chinese-hubert-base.onnx"

# Optional: point to the Open JTalk dictionary. If omitted, the script will try to download it from GitHub.
os.environ['OPEN_JTALK_DICT_DIR'] = r"C:\path\to\your\open_jtalk_dic_utf_8-1.11"

import lunavox_tts as lunavox

# Step 1: load the character ONNX bundle
lunavox.load_character(
    character_name='<CHARACTER_NAME>',
    onnx_model_dir=r"<PATH_TO_CHARACTER_ONNX_MODEL_DIR>",
)

# Step 2: set the reference audio (voice cloning prompt)
lunavox.set_reference_audio(
    character_name='<CHARACTER_NAME>',
    audio_path=r"<PATH_TO_REFERENCE_AUDIO>",
    audio_text="<REFERENCE_AUDIO_TEXT>",
    audio_language='ja',  # ja / zh / en
)

# Step 3: synthesise speech
lunavox.tts(
    character_name='<CHARACTER_NAME>',
    text="<TEXT_TO_SYNTHESIZE>",
    play=True,
    save_path="<OUTPUT_AUDIO_PATH>",
    language='ja',  # Target language
)

print("Audio generated.")

The following numbers were collected with benchmark/scripts/tts_benchmark.py on Windows 11, Python 3.12, 32 GB RAM, and an Intel Core i9-12900K. Each run used 3 warm-up iterations plus 100 measured loops with the fixed text “This is LunaVox speaking English.”

• Both models achieve a real-time factor of roughly 0.54, producing audio faster than real time.
• Full metrics and per-iteration logs are stored in benchmark/results/v2_results.json and benchmark/results/v2_pro_plus_results.json.

Install the optional converter dependencies first:

pip install "lunavox-tts[convert]"

import lunavox_tts as lunavox

lunavox.convert_to_onnx(
    torch_pth_path=r"<PATH_TO_PTH>",
    torch_ckpt_path=r"<PATH_TO_CKPT>",
    output_dir=r"<OUTPUT_ONNX_DIR>",
)

The converter decomposes the GPT-SoVITS pipeline into multiple ONNX graphs: t2s_encoder_fp32.onnx, t2s_first_stage_decoder_fp32.onnx, t2s_stage_decoder_fp32.onnx, and vits_fp32.onnx, while bundling the Chinese HuBERT model and speaker vector network. During conversion the original FP16 weights are temporarily promoted to FP32 so that ONNX Runtime delivers stable numerical behavior on CPU-only hosts.

• LUNAVOX_ORT_PROVIDERS: override the preferred ONNX Runtime providers (comma-separated). Example: CUDAExecutionProvider,CPUExecutionProvider.
• LUNAVOX_USE_IO_BINDING=1: enable experimental IO binding for the vocoder step (can reduce host/device copies when GPU providers are available).

import os

os.environ['HUBERT_MODEL_PATH'] = r"C:\path\to\your\chinese-hubert-base.onnx"
os.environ['OPEN_JTALK_DICT_DIR'] = r"C:\path\to\your\open_jtalk_dic_utf_8-1.11"

import lunavox_tts as lunavox

lunavox.start_server(
    host="0.0.0.0",
    port=8000,
    workers=1,
)

See Tutorial/English/API Server Tutorial.py for request formats and endpoint details.

LunaVox includes a Gradio-based web interface for browser-based synthesis.

# Windows
start_webui.bat

# Or run directly
python WebUI/webui.py

• Character management: automatically scans Data/character_model
• Reference audio: upload custom prompts or reuse the included samples
• Text synthesis: enter Japanese text and generate speech with one click
• In-browser playback: listen instantly within the UI
• File saving: generated audio is saved under Output

1. After launching, the browser opens http://127.0.0.1:7860
2. Select a character model (the ONNX bundle loads automatically)
3. Provide a reference audio clip (upload or choose from presets)
4. Enter the text to synthesise
5. Click “Generate” to produce and preview the audio

import lunavox_tts as lunavox

lunavox.launch_command_line_client()

• Language expansion

  • Chinese support
  • English support

• Model compatibility

  • GPT-SoVITS V2 Pro support
  • GPT-SoVITS V2 Pro Plus support

• Performance improvements

  • Publish a GPU-oriented build
  • Implement text-splitting utilities for long-form synthesis

• Easier deployment

  • Publish a Docker image
  • Provide ready-to-use Windows / Linux bundles