轻量级本地 TTS Kokoro-82M 完整部署教程

# 创建虚拟环境
python -m venv kokoro-env

# Windows激活环境
kokoro-env\Scripts\activate
# Linux/Mac激活环境
source kokoro-env/bin/activate

# 主模型包 + 中文分词音素库（必装）
pip install kokoro misaki[zh] soundfile

# 安装下载工具
pip install huggingface-hub
# 拉取中文模型到本地文件夹
huggingface-cli download hexgrad/Kokoro-82M-v1.1-zh --local-dir ./kokoro-zh

from kokoro import KPipeline
import soundfile as sf

# 初始化中文管线 lang_code='z'=中文
pipeline = KPipeline(lang_code='zh')

# 待合成文本
text = "关注公众号奥德元，一起学习AI，一起追赶时代。"
# 选择音色 zm_yunyang浑厚男声 / zf_xiaoxiao成熟女声
generator = pipeline(text, voice="zm_yunyang")

# 遍历生成分段音频并保存
for idx, (grapheme, phoneme, audio_data) in enumerate(generator):
    sf.write(f"output_{idx}.wav", audio_data, 24000)
print("语音合成完成，文件已保存！")

from kokoro import KPipeline
import soundfile as sf
import numpy as np
import torch

def tts_test():
    # 初始化中文管道
    pipeline = KPipeline(lang_code='zh')
    text = "今天天气晴转多云，最高25度，最低18度，适合出门。"
    generator = pipeline(text, voice='zf_xiaoyi')
    result = next(generator)
    
    # 提取音频张量并转换格式
    audio_tensor = result.output.audio
    audio_numpy = audio_tensor.detach().cpu().numpy()
    
    # 统一维度适配soundfile
    if audio_numpy.ndim == 1:
        audio_numpy = audio_numpy.reshape(-1, 1)
    
    sample_rate = 24000
    sf.write("weather.wav", audio_numpy, sample_rate)
    
    print(f"音频保存成功！采样率：{sample_rate}Hz")
    print(f"文本：{result.graphemes}")

if __name__ == "__main__":
    tts_test()

from kokoro import KPipeline
import soundfile as sf
import os

pipeline = KPipeline(lang_code='zh')

def speak_weather():
    text = "今日天气晴转多云，最高温度25度，最低温度18度，微风。"
    gen = pipeline(text, voice='zf_xiaoyi')
    for _, _, audio in gen:
        sf.write("weather_notice.wav", audio, 24000)
    # Linux播放音频
    os.system("aplay weather_notice.wav")

speak_weather()

from kokoro import KPipeline
import soundfile as sf
import os

pipeline = KPipeline(lang_code='zh')

def generate_audiobook(txt_path, out_dir, voice="zf_xiaoxiao"):
    os.makedirs(out_dir, exist_ok=True)
    with open(txt_path, "r", encoding="utf-8") as f:
        full_text = f.read()
    # 按段落拆分，防止长文本内存溢出
    paragraphs = full_text.split("\n")
    for p_idx, para in enumerate(paragraphs):
        if not para.strip():
            continue
        gen = pipeline(para, voice=voice)
        for seg_idx, (_, _, audio) in enumerate(gen):
            save_path = f"{out_dir}/chapter_{p_idx:03d}_{seg_idx}.wav"
            sf.write(save_path, audio, 24000)
        print(f"已完成段落 {p_idx+1}/{len(paragraphs)}")

# 调用，将novel.txt批量转为音频
generate_audiobook("novel.txt", "audio_book_output")

# 中文
zh_pipe = KPipeline(lang_code="zh")
zh_audio = zh_pipe("你好，世界", voice="zf_xiaoxiao")

# 英语
en_pipe = KPipeline(lang_code="en")
en_audio = en_pipe("Hello World", voice="af_sarah")

# 日语（需额外安装misaki[ja]）
# pip install misaki[ja]
ja_pipe = KPipeline(lang_code="ja")
ja_audio = ja_pipe("こんにちは世界", voice="jf_sarah")