moss-tts-local-onnx-cli 0.1.0

MOSS-TTS-Local-Transformer-v1.5 voice cloning via ONNX Runtime

moss-tts-local-onnx-cli

MOSS-TTS-Local-Transformer-v1.5 voice cloning via ONNX Runtime. Converted from the original PyTorch model (4.13B params, Qwen3 global + GPT2 local transformer).

Quick Start

One-liner with uvx (auto-downloads ~31GB models on first run):

uvx moss-tts-local-onnx-cli --text "你好，这是一段测试语音合成。" --reference ref.wav --output output.wav

Installation

# Install with uv
uv pip install moss-tts-local-onnx-cli

# Or with pip
pip install moss-tts-local-onnx-cli

# For ModelScope download support (China mainland):
pip install "moss-tts-local-onnx-cli[modelscope]"

CLI Usage

# Basic voice cloning
moss-tts-local-onnx-cli --text "你好" --reference speaker.wav --output output.wav

# With language hint
moss-tts-local-onnx-cli --text "Hello world" --reference speaker.wav --language English --output output.wav

# Download from ModelScope (China)
moss-tts-local-onnx-cli --text "你好" --reference speaker.wav --source modelscope --output output.wav

# Custom model directory
moss-tts-local-onnx-cli --text "你好" --reference speaker.wav --model-dir /path/to/models --output output.wav

Python API

from moss_tts_local_onnx_cli import MossTTS

# Initialize once (auto-downloads models)
tts = MossTTS()

# Generate multiple times without reloading
audio = tts.generate("你好，这是第一段。", reference="ref.wav", language="Chinese")
tts.save(audio, "output1.wav")

audio2 = tts.generate("这是第二段。", reference="ref.wav")
tts.save(audio2, "output2.wav")

# Or generate and save in one call
tts.generate("你好", reference="ref.wav", output="output3.wav")

Audio Codec

The audio codec (MOSS-Audio-Tokenizer-v2) is required for decoding audio tokens. It will be auto-detected from ~/.cache/modelscope/hub/models/OpenMOSS/MOSS-Audio-Tokenizer-v2.

To download it:

pip install modelscope
python -c "from modelscope import snapshot_download; snapshot_download('OpenMOSS/MOSS-Audio-Tokenizer-v2')"

Or specify a custom path:

tts = MossTTS(codec_path="/path/to/MOSS-Audio-Tokenizer-v2")

Model Conversion

To convert the original PyTorch model to ONNX yourself:

# Install export dependencies
pip install torch transformers onnx

# Run export (requires ~18GB RAM)
python scripts/export_onnx.py

# Validate correctness
python scripts/validate.py

See scripts/export_onnx.py for the full conversion pipeline.

Model Architecture

ONNX Model	Description	Size
global_prefill.onnx	Qwen3 36-layer full-sequence forward	~15GB
global_decode.onnx	Qwen3 single-token decode with KV cache	~15GB
local_transformer_first.onnx	GPT2 1-layer (first step)	290MB
local_transformer.onnx	GPT2 1-layer (with KV cache)	290MB
lm_heads.npz	LM head weight matrices	120MB

Requirements

• Python >= 3.10
• ~32GB RAM (for loading models)
• ~31GB disk space (for model files)

License

MIT