piper-voices-zh-AISHELL-3 (cn_aishell3, medium)

A standard Piper multi-speaker Mandarin (zh_CN) voice, trained on AISHELL-3.

• Language: zh_CN (Chinese, China)
• Frontend: pinyin (g2pW polyphone disambiguation) — phoneme_type=pinyin
• Speakers: 174 (ids/names are the original AISHELL-3 speaker codes, e.g. SSB0033)
• Quality: medium (22,050 Hz)
• This is a standard Piper ONNX voice.

Note: espeak is not used at inference for Chinese. Piper's pinyin frontend downloads the g2pW model automatically (from rhasspy/piper-checkpoints zh/zh_CN/_resources/g2pw.tar.gz) on first use.

Files

zh/zh_CN/cn_aishell3/medium/
  zh_CN-cn_aishell3-medium.onnx
  zh_CN-cn_aishell3-medium.onnx.json

Training checkpoint

Checkpoint + config: https://huggingface.co/datasets/writeAPoemForYou/piper-checkpoints-zh-AISHELL-3

Install

pip install piper-tts          # provides `python -m piper`

Download this voice

pip install -U "huggingface_hub[cli]"
hf download writeAPoemForYou/piper-voices-zh-AISHELL-3 \
  zh/zh_CN/cn_aishell3/medium/zh_CN-cn_aishell3-medium.onnx \
  zh/zh_CN/cn_aishell3/medium/zh_CN-cn_aishell3-medium.onnx.json \
  --local-dir cn_aishell3

Inference (CLI)

python -m piper \
  --model cn_aishell3/zh_CN-cn_aishell3-medium.onnx \
  --speaker SSB0033 \
  -f out.wav \
  -i "这是一本书。"

Inference (Python, verified)

import wave
from pathlib import Path
import numpy as np
from piper import PiperVoice
from piper.config import SynthesisConfig

MODEL_DIR = Path("cn_aishell3")
voice = PiperVoice.load(
    MODEL_DIR / "zh_CN-cn_aishell3-medium.onnx",
    download_dir=MODEL_DIR,          # where g2pW is cached/downloaded
)

# multi-speaker: pick a speaker (0..173, or the name in speaker_id_map)
syn = SynthesisConfig(speaker_id=0)  # 0 == "SSB0033"

chunks = list(voice.synthesize("这是一本书。", syn_config=syn))
audio = np.concatenate([c.audio_float_array for c in chunks])
sr = chunks[0].sample_rate

with wave.open("out.wav", "wb") as w:
    w.setnchannels(1)
    w.setsampwidth(2)
    w.setframerate(sr)
    w.writeframes((np.clip(audio, -1, 1) * 32767).astype(np.int16).tobytes())
print("wrote out.wav", audio.shape, sr)

Speakers

zh_CN-cn_aishell3-medium.onnx.json → speaker_id_map maps the original AISHELL-3 speaker codes (e.g. SSB0033) to ids 0..173.

Dataset / License

• Dataset: AISHELL-3 — https://www.openslr.org/93/
• License: Apache License v.2.0
• Warm-started from en_US/libritts_r (medium) and fine-tuned on AISHELL-3.

@inproceedings{AISHELL-3_2020,
  title={AISHELL-3: A Multi-speaker Mandarin TTS Corpus and the Baselines},
  author={Yao Shi, Hui Bu, Xin Xu, Shaoji Zhang, Ming Li},
  year={2015},
  url={https://arxiv.org/abs/2010.11567}
}