ayousanz / piper-plus-moe

piper-plus-moe

Japanese Multi-Speaker Text-to-Speech Model based on VITS architecture, compatible with Piper TTS .

Model Overview

Usage

With Piper TTS

# Download model files
wget https://huggingface.co/ayousanz/piper-plus-moe/resolve/main/moe-speech-20speakers-200epochs.onnx
wget https://huggingface.co/ayousanz/piper-plus-moe/resolve/main/moe-speech-20speakers-200epochs.onnx.json

# Run inference with Piper
echo "こんにちは、私は日本語音声合成モデルです。" | piper \
  --model moe-speech-20speakers-200epochs.onnx \
  --output_file output.wav \
  --speaker 0

Training Command

NCCL_DEBUG=INFO NCCL_P2P_DISABLE=1 NCCL_IB_DISABLE=1 uv run python -m piper_train \
  --dataset-dir /data/piper/dataset-moe-speech-20speakers \
  --accelerator gpu --devices 4 --precision 16-mixed \
  --max_epochs 200 --batch-size 32 --samples-per-speaker 4 \
  --checkpoint-epochs 1 --quality medium \
  --base_lr 2e-4 --disable_auto_lr_scaling \
  --ema-decay 0.9995 --num-workers 0 --no-pin-memory \
  --default_root_dir /data/piper/output-moe-speech-20speakers-lr2e4-fixed

Dataset

• Source : moe-speech dataset
• Total Speakers : 20
• Total Utterances : ~60,000
• Total Duration : ~90 hours
• Phoneme Type : OpenJTalk

Citation

@misc{piper-plus-moe,
  author = {ayousanz},
  title = {piper-plus-moe: Japanese Multi-Speaker TTS Model},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/ayousanz/piper-plus-moe}
}

Acknowledgements

• Piper TTS - Original TTS framework
• VITS - Model architecture
• moe-speech dataset contributors