moss-tts-local-mnn-cli 0.1.0

MOSS-TTS-Local-Transformer-v1.5 voice cloning via MNN

moss-tts-local-mnn-cli

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

MNN provides optimized CPU inference on ARM (Apple Silicon, Android) with fp16/int8 quantization support.

Quick Start

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

Installation

uv pip install moss-tts-local-mnn-cli

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

CLI Usage

# Voice cloning with fp16 (default)
moss-tts-local-mnn-cli --text "你好" --reference speaker.wav --output output.wav

# Use int8 quantization (smaller, faster)
moss-tts-local-mnn-cli --text "你好" --reference speaker.wav --quantization int8 --output output.wav

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

Python API

from moss_tts_local_mnn_cli import MossTTS

# Initialize once
tts = MossTTS(quantization="fp16")

# 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")

Audio Codec

The audio codec (MOSS-Audio-Tokenizer-v2) is required for decoding. Download it:

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

Model Conversion

Convert ONNX models to MNN:

# Prerequisites: build MNNConvert from https://github.com/alibaba/MNN

# fp16
MNNConvert -f ONNX --modelFile global_prefill.onnx --MNNModel global_prefill_fp16.mnn --fp16 --saveExternalData
MNNConvert -f ONNX --modelFile global_decode.onnx --MNNModel global_decode_fp16.mnn --fp16 --saveExternalData
MNNConvert -f ONNX --modelFile local_transformer_first.onnx --MNNModel local_transformer_first_fp16.mnn --fp16
MNNConvert -f ONNX --modelFile local_transformer.onnx --MNNModel local_transformer_fp16.mnn --fp16

# int8
MNNConvert -f ONNX --modelFile global_prefill.onnx --MNNModel global_prefill_int8.mnn --weightQuantBits 8 --saveExternalData
MNNConvert -f ONNX --modelFile global_decode.onnx --MNNModel global_decode_int8.mnn --weightQuantBits 8 --saveExternalData
MNNConvert -f ONNX --modelFile local_transformer_first.onnx --MNNModel local_transformer_first_int8.mnn --weightQuantBits 8
MNNConvert -f ONNX --modelFile local_transformer.onnx --MNNModel local_transformer_int8.mnn --weightQuantBits 8

Benchmark (Apple M3 Max, 36GB)

Engine	ms/step	vs PyTorch
PyTorch fp32	338	1.00x
ONNX fp32	374	0.90x
MNN fp16	274	1.24x
MNN int8	277	1.22x

Requirements

• Python >= 3.10
• ~20GB RAM (fp16) or ~16GB RAM (int8)
• ~14GB disk (fp16) or ~11GB disk (int8)

License

MIT